acceptodds
Under review as a conference paper at ICLR 2027

HIDDENFACE: STREAMING CONVERSATIONAL FACIAL MOTION FROM HIDDEN STATES OF A LARGE AUDIOLANGUAGE MODEL

17 views

Abstract

Full-duplex large audio-language models (LALMs) enable voice agents to take part in conversations, but embodying these agents as avatars also requires facial motion that stays aligned with their speech and continues through listening and turn transitions. However, audio-driven methods must reprocess the generated speech in a cascade, incurring processing and synchronization overhead. Methods that drive facial motion from LALM internal states mitigate this, but they train the LALM jointly, rely on additional generated speech units, or delay the output to refine motion with later hidden states. We propose HiddenFace, which generates conversational facial motion using only the hidden states already computed by a frozen LALM (Voila) during speech generation. HiddenFace combines Voila's audio and text hidden states with its own motion-token history to autoregressively predict the next facial motion token, finalizing each 160 ms block as soon as its hidden states are available, without waiting for future speech. Under matched 160 ms block-causal conditions, HiddenFace matches or exceeds the lip accuracy of audio-driven baselines on held-out real dyadic conversations and adds only a small, constant per-step cost to speech generation. At the dialogue level, it is closer to real human facial behavior than any baseline, even those using the full audio, and receives the highest mean ratings in a human study.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Large Audio-Language Model, Conversational Facial Motion Generation, Audio–Text Hidden-State Fusion, Autoregressive Motion Generation, Full-Duplex Voice Agent

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_X1AhOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

HiddenFace predicts facial-motion codes every 160 ms from the audio and text hidden states of a frozen large audio-language model. A causal RVQ motion tokenizer, a short previous-segment memory and a depth-wise code predictor support streaming generation. The paper also develops geometric surrogate objectives and uses synthetic self-play conversations for training. It evaluates real conversations, generated conversations, human preferences and runtime.

Critical strengths

The motion head is small and fast, and attaching it to a frozen conversational model is useful. The paper gives substantial implementation detail, compares hidden states with audio codes, studies auxiliary losses and memory, and reports limitations and metric trade-offs. The human study uses blinded presentation and rater bootstrap intervals. The authors also discuss closely related systems, including Ex-Omni, FacePlex and DyaPlex, rather than presenting speech–motion generation as a new problem.

Critical weaknesses

1. Causality in the audio-language model's tokens is not sufficient to establish causality in the input waveform. Sec. 4.2 describes the streaming comparison as using no audio beyond the current 160 ms block. However, Appendix C says each real recording is encoded once with the non-streaming Voila tokenizer, and prefix re-encoding shows that a 40 ms code frame can depend on up to four later frames, with a median of three. A causal transformer operating on these codes can therefore inherit future waveform information.

The key missing detail is whether the real-conversation test states in Table 1 are extracted using this same full-recording procedure. If so, the claimed matched causal comparison is not established: the inputs to HiddenFace may contain future information even when its motion predictor is causal. Please rerun with prefix-only audio tokenization, or measure the actual waveform lookahead and give the baselines the same allowance. Appendix F's past-only re-encoding experiment shows one possible live implementation, but does not establish that the main quantitative comparison used the same input boundary. This is a question about the evaluation protocol; the appendix alone does not prove that every reported result used future audio.

2. The practical and comparative evidence is incomplete for the claimed streaming interface. The 5.8 ms number measures the motion component with CUDA graphs. Appendix F gives a more informative system experiment, but it uses separate LLM/codec GPUs, excludes text-generation stalls of up to 1.6 seconds from the real-time subset, and reports its steady-state table after decoder warm-up. These choices are described, but the resulting timing should not stand in for the latency a user experiences. Please report all-block end-to-end latency, including startup and speech/text stalls, with tail percentiles, missed-deadline rates and interruption examples.

The main motion baselines are also adapted offline systems. The paper already cites closer conversational systems, but there is no quantitative comparison with a streaming speech–motion architecture. In particular, DyaPlex [1] also connects a trainable motion pathway to hidden states of a frozen speech model, and FacePlex [2] is designed for online joint speech/facial motion. Their backbones and tasks differ, so an unadapted score would not be a fair comparison; a matched-backbone implementation or a carefully controlled component comparison would be more useful. The audio-code ablation is informative, but does not replace that comparison. The human study does not establish a significant advantage over DualTalk, as the authors themselves acknowledge.

The system is promising. I currently lean toward rejection because the main causal comparison needs to be resolved, and the deployment evidence does not yet cover the whole streaming system.

Questions for authors

Questions For Authors

  1. Were Table 1's real-test hidden states extracted from full-recording audio codes? Please specify the waveform context available to every code/state in the current block.
  2. What happens to quality under prefix-only tokenization, with any lookahead charged to the latency budget? Can all methods use the same waveform boundary?
  3. Please report end-to-end startup and steady-state p50/p95/p99 latency across all blocks, including text stalls, plus deadline misses and interruption behavior.
  4. Can you compare against a closely related streaming speech–motion architecture under a matched backbone/data budget, separating the benefit of hidden-state input from the predictor design?

References

[1] DyaPlex: Full-Duplex Speech-Motion Model for Dyadic Interaction. Secs. 3–4 describe a frozen speech tower and a trainable motion tower attending to its hidden states. https://arxiv.org/html/2606.03874v1

[2] FacePlex: Full-Duplex Joint Speech-Facial Motion Generation for Conversational Avatars. https://arxiv.org/html/2606.30145v1

7 comments
AuthorsOpenReview ↗

Response to Reviewer X1Ah (1/6): W1, Q1, Q2

Thank you for the specific questions about the evaluation protocol. We have uploaded a revised paper with the changes marked in blue. The sections, tables, and appendices cited below refer to this revision, and Appendices H, I, and J are new.

We first summarize our answers to the two reasons you gave for leaning toward rejection.

First, the states in Table 1 were extracted from codes obtained by encoding each recording once. As you suggested, we repeated training and evaluation with prefix-only tokenization. When only the partner's audio is encoded from prefixes, the comparison with each released baseline of Table 1 gives the same result as before on all 12 metrics. We explain under W1 why the agent's audio stays encoded once in this setting. When both speakers are encoded from prefixes, so that all methods use the same waveform boundary, the model retrained in this setting remains better than streaming UniLS on eight of the 12 metrics but is less accurate in S-LVE and MOD (W1, Q1, Q2).

Second, we now report every block of the live run, including the startup and the text-generation stalls. Of the 999 blocks after the first, 135 (13.5%) were emitted more than 160 ms after they ended, and the 99th-percentile latency is 1.30 s. No block missed the 160 ms deadline from waiting for the face (W2, Q3).

For the comparison with related architectures, we change one core design choice at a time on the same backbone and data (W2, Q4).

W1, Q1, Q2 (waveform causality of the Table 1 inputs)
"Were Table 1's real-test hidden states extracted from full-recording audio codes? Please specify the waveform context available to every code/state in the current block. What happens to quality under prefix-only tokenization, with any lookahead charged to the latency budget? Can all methods use the same waveform boundary?"

Yes. The test states in Table 1 were extracted from codes obtained by encoding each recording once ("encoded once" in the tables), in the same way as the training states. Each state is computed only from the tokens before its row and the predictor reads states only up to the current block, so HiddenFace is causal at the token level (Appendix C). As you point out, however, the released Voila [3] tokenizer is not a streaming codec, and a 40 ms code frame depends on up to four later audio frames (three at the median). Table A lists how much later audio each input of Voila and the resulting states can contain under three tokenization settings.

Waveform context at row ttEncoded once / encoded oncePrefix-only / encoded oncePrefix-only / prefix-only
Partner audio code160 ms past row tt*to end of row ttto end of row tt
Agent audio code160 ms past row tt*160 ms past row tt*to end of row tt
Agent transcript wordfrom its onset row (Sec. 3.4)samesame
Voila states AtA_t, TtT_tfrom codes and words up to row t−1t-1samesame
All inputs of the block120 ms past block end*partner 0, agent 120 ms*0 (as the baselines)

Table A. Waveform context of the inputs of Voila and of the resulting states at row tt (partner / agent). * Measured dependence, excluding the rare isolated code changes noted in Appendix C. These values are not verified worst-case bounds.

In our measurement, a code frame depends on up to 160 ms of later audio. Because the states AtA_t and TtT_t are computed from the codes up to row t−1t-1, the last states of a block then reflect up to 120 ms of audio beyond the block.

Whether this look-ahead departs from real-time operation depends on the speaker. In real-time operation, the partner's audio is encoded from only the audio received so far. Partner codes encoded once thus contain audio beyond the end of the block that a real-time system cannot know. The agent's codes are instead generated by Voila and not encoded from a waveform, so they involve no tokenizer look-ahead and no waiting that would be charged to the latency budget.

Real conversations, however, contain no codes generated by Voila. Proxy codes encoded from the recorded audio are therefore needed, and they should be close to the codes Voila generates. Prefix-only tokenization encodes each code frame from the waveform up to the end of that frame. When speech generated by Voila is re-encoded and compared with the generated codes, all four codebooks match in 24.5% of the speech frames (86.9% for the first codebook) if the audio is encoded once but in only 3.0% (57.1%) with prefix-only tokenization. By this measure, encoding once is the closer of the two proxies for the agent.

AuthorsOpenReview ↗

Response to Reviewer X1Ah (2/6): W1, Q1, Q2, continued

We followed the first of your two suggestions. We applied prefix-only tokenization in two settings and re-extracted the Voila states of both the training and test sets. In the partner-prefix setting, only the partner's audio is encoded from prefixes, and the agent's codes remain encoded once as the closer proxy for the codes Voila generates. In the both-prefix setting, both speakers are encoded from prefixes, so that every input reflects only the audio up to the end of the current block. In each setting, we evaluated the submitted checkpoint as is ("submitted") and a model retrained with the same code, hyperparameters, seed, and training schedule ("retrained").

Tokenization (partner / agent)ModelS-LVE↓S-MHD↓MOD↓Lip-W↓S-FDD↓L-FDD↓AVC↑MTM↓MSE↓FD↓P-FD↓rPCC↓
UniLS [4] (released), 160 ms streaming (Table 1)–8.6042.4843.7253.94024.5822.990.30558.40.24115.3216.440.135
Prefix-only / prefix-onlysubmitted8.4262.2903.7973.55325.3327.090.26681.70.18511.7612.750.114
Prefix-only / prefix-onlyretrained8.7322.3383.9293.47423.2926.500.32774.60.18811.6712.660.116
Encoded once / encoded oncesubmitted (Table 1)8.2292.2473.7303.45223.3326.880.37463.90.17811.0011.960.113
Prefix-only / encoded oncesubmitted8.1982.2423.7513.50123.1826.640.37762.40.17810.9511.920.112
Prefix-only / encoded onceretrained (Table 1)8.1892.2443.7113.45923.3226.520.38861.30.17910.9411.910.112

Table B. Results under the three tokenization settings (partner / agent).

For each of the 12 metrics, we compare HiddenFace with each of the seven released baseline rows of Table 1. For each comparison, we compute the difference between HiddenFace and the baseline on the same test items and obtain its 95% confidence interval by a paired bootstrap over the 1,554 continuous runs of the test split. The comparison counts as not different if this interval includes 0 and as better or worse otherwise.

In the partner-prefix setting, every comparison gives the same result as in Table 1 with or without retraining. For the submitted checkpoint without retraining, every metric changes by at most 2.4%.

In the both-prefix setting, every input ends at the end of the block (last row of Table A). This is the same boundary as the streaming input of the baselines (Appendix E: UniLS and ARTalk set the audio not yet received to zero, and DualTalk is rerun on the conversation up to each step). No look-ahead beyond the 160 ms block therefore remains, and charging the look-ahead to the latency budget adds no delay.

Table B lists the both-prefix setting for the submitted checkpoint without retraining and for a model retrained in it. The retrained both-prefix model remains the best among the released streaming methods on eight of the 12 metrics, but it is less accurate than streaming UniLS in S-LVE and MOD. Four of its advantages over offline baselines are also no longer significant. The other comparisons give the same result as in Table 1. Because the agent's codes in this setting differ greatly from those Voila generates, we regard it as a waveform-causal replay rather than as the condition of real-time operation.

The live run of Appendix F indeed does not establish the input boundary of Table 1. Table B therefore does not rely on Appendix F. It repeats the quantitative comparison itself with prefix-only inputs.

The remaining HiddenFace results on real conversations (Table 1 except its partner-prefix rows, and the ablations in Tables 3, 4, and 7) use states extracted from codes encoded once. The baseline rows and the latency columns of Table 1 do not pass through the Voila tokenizer and are unaffected. The self-play conversations of the human study (Table 2) do not pass through it either, because the two Voila instances exchange the codes they generate. Only the audio that opens a conversation is encoded once.

Sec. 4.1 of the revision now states that the hidden states of Table 1 can reflect audio beyond the block, up to 120 ms in our measurement. It also reports the partner-prefix result. The both-prefix result is in Sec. 4.2 and Appendix H, and the waveform context of each setting is in Appendix H. Table 1 has new partner-prefix rows, so the partner-prefix result can be read from the main table.

We also corrected the wording you pointed out. Where the paper called the comparison "matched" or "same condition" (abstract, Sec. 1, the sub-header of Table 1, Sec. 4.2, and Sec. 5), we removed these words and now write only "block-causal".

AuthorsOpenReview ↗

Response to Reviewer X1Ah (3/6): W2 (first part), Q3

W2 (first part), Q3 (latency over all blocks and interruptions)
"Please report end-to-end startup and steady-state p50/p95/p99 latency across all blocks, including text stalls, plus deadline misses and interruption behavior."

We address the three choices you list in turn (the separate LLM/codec GPUs, the excluded text-generation stalls, and the steady state reported after decoder warm-up). We use two GPUs because the system did not run in real time on one. In a single-GPU run, the computation of Voila's speech path alone took 1.06 times real time, and the facial branch added another 0.22. In the two-GPU setup, Voila's LLM and audio tokenizer run on separate GPUs, and the facial branch adds 6.3 ms per block on the tokenizer GPU. We have no run of Voila without the facial branch, so these measurements do not show how much sharing a GPU with it slows Voila itself.

For the other two choices, we now analyze every block. This includes the intervals in which Voila generates text tokens in succession and its audio pauses, the intervals in which it then processes the queued input, and the interval right after the run starts. We leave out only the first block, which contains Voila's initialization, and analyze the remaining 999 blocks.

BlocksNFace step, med / max (ms)Block-end-to-emission p50 / p95 / p99 (ms)Late > 160 ms (from waiting for face)Delayed by face (max)
Steady state7966.3 / 11.49.4 / 19.3 / 19.60 (0)1 (2.1 ms)
+ startup8496.3 / 28.19.4 / 20.3 / 30.80 (0)35 (16.9 ms)
+ text stalls, catch-up (all)9996.3 / 28.19.5 / 881 / 1300135 (0)35 (16.9 ms)

Table C. Latency over the 999 blocks after the first block of the live run.

Step in Table 1 (5.8 ms) is the time the facial branch takes to compute the motion of one block, measured with the facial branch alone. Face step in Table C is the same computation measured while running with Voila. Block-end-to-emission, in contrast, is not a computation time. It is the latency of Voila and the facial branch together, from the end of a block's fourth row interval until both its audio and face are ready and the block is emitted. In the steady state, the face is ready 5.0 ms after the block ends and the fourth row's audio becomes playable at 9.4 ms, so the 9.4 ms in Table C is the time Voila takes to prepare the audio. Because audio and motion are emitted per block, the first row of a block is emitted about 130 ms after its own interval ends. This wait comes from emitting motion in 160 ms blocks and not from waiting for future context.

We distinguish the blocks that waited for the face from those whose audio was already late. Of the 999 blocks, 135 (13.5%) were emitted more than 160 ms after they ended. All of them fall in the intervals in which Voila generates text tokens in succession (1.59, 1.13, and 0.88 s) and then processes the queued partner input. In these blocks Voila's own audio is late, and no block missed the 160 ms deadline from waiting for the face. This is also why the 95th and 99th percentiles over all blocks rise to 0.9–1.3 s.

Separately, the face is usually ready before the audio and waits for it. The face was ready after the audio in only 35 of the 999 blocks, so that the block waited for the face (Delayed by face in Table C). The wait was at most 16.9 ms, and these 35 blocks were all emitted within 160 ms. Of them, 34 fall in the first 53 blocks. During these blocks the motion decoder's input window has not yet filled and CUDA graphs cannot be applied, so the face step takes 22 ms instead of 6.3 ms.

Adding the startup changes the latency only slightly (99th percentile: 30.8 ms). The first motion is emitted 0.21 s after the run starts (0.16 s to receive the input of the first block, 0.02 s for Voila to process its last row, and 0.03 s for the first face computation).

We added this table and analysis to Appendix F, and Sec. 4.2 now reports its main numbers.

For interruptions, we collected all cases in which the partner starts speaking while the agent is speaking in three live runs and found 26, 26, and 23 cases (Appendix F). Voila decides whether the agent then stops or continues, and the face follows Voila's hidden states. We therefore examined whether mouth motion also subsides once Voila stops speaking. Because the partner's speech is a recording prepared in advance and does not react to the agent, these runs do not test interaction, and we do not claim that the system handles interruptions.

In these cases Voila mostly keeps speaking (it stops within 2 s in 15–42% of them). When it stops, mouth motion subsides within one to two blocks. The speed at which the mouth opens and closes averages 34–48 mm/s in the 0.4 s before the stop, 8–13 mm/s in the first 0.2 s after it, and 2–3 mm/s in the next 0.2 s. In one case of the Appendix F run, the partner starts speaking at 28.52 s, Voila stops 0.16 s later, and the speed falls from 19.9 to 2.1 mm/s within 0.2 s of the stop.

Reviewer_PMQEOpenReview ↗
Rating
3
Confidence
4
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

HiddenFace generates conversational facial motion directly from a frozen audio-language model’s hidden states, supporting speaking, listening, and turn transitions without reprocessing audio or waiting for future speech. It achieves good lip accuracy compared with audio-driven baselines with little computational overhead, while producing more human-like dialogue behavior.

Critical strengths
  • This paper aims to eliminate the need to reprocess generated audio and reduce overall system latency, which is important for conversational avatar applications.
  • The fully causal design also helps keep latency bounded.
  • Approximating vertex velocities in latent space is a clever design.
Critical weaknesses
  • The paper lacks direct comparisons with closely related methods such as A2-LLM and FacePlex.
  • The Voila backbone should provide access to semantic information, but the paper does not directly evaluate whether the generated reactions are semantically appropriate.
Questions for authors
  • Compared with A2-LLM, this paper keeps the backbone frozen. However, if joint fine-tuning does not noticeably degrade the backbone’s capabilities, it seems potentially more effective. Have the authors tried jointly fine-tuning Voila? Are the benefits of freezing primarily computational, or are there also differences in performance?
  • Is HiddenFace’s main facial-motion branch deterministic? The paper appears to use greedy decoding for token selection. How would introducing randomness through sampling affect the results?
  • Direct comparisons with A2-LLM, Moshi-Face, and FacePlex may be difficult because of differences in backbones and potential code availability issues. Could ablations approximate some of their core design choices—for example, introducing non-causal decoding or future context—to examine the effects on performance? It is entirely understandable if these experiments are infeasible during rebuttal.
  • The method appears to update its history memory at segment boundaries. Does this mean that predictions within a segment can access the previous segment’s motion tokens, but not tokens already predicted within the current segment? Is this interpretation correct, and could this design cause discontinuities?
  • Could the authors report the overall end-to-end latency? This would help assess the efficiency of the complete system.
5 comments
AuthorsOpenReview ↗

Response to Reviewer PMQE (1/4): W1, Q1, Q3

Thank you for the positive assessment and the specific questions. We have uploaded a revised paper with the changes marked in blue. The sections, tables, and appendices cited below refer to this revision, and Appendices H, I, and J are new.

Before the answers, we describe the setting behind the comparisons in Table A. In Table 1, the audio of the real conversations was encoded once per recording, and in our measurement a code frame of the released Voila [1] tokenizer depends on up to 160 ms of later audio. We therefore re-extracted the hidden states with the partner's audio encoded from only the part received so far (partner-prefix) and repeated training and evaluation. For the agent, we kept the codes encoded once from the recording as the proxy for the codes Voila generates in real-time operation. Appendix H also reports the setting in which both speakers are encoded from prefixes.

In the partner-prefix setting, the values of HiddenFace stay close to those of Table 1 (S-LVE: 8.229 vs. 8.189; AVC: 0.374 vs. 0.388). The comparison with each baseline also gives the same result as in Table 1 on all 12 metrics (Appendix H). We added the partner-prefix model as new rows in Table 1, and all rows of Table A are trained and evaluated in this setting.

W1, Q1, Q3 (comparison with A²-LLM, FacePlex, and Moshi-Face)
"Could ablations approximate some of their core design choices—for example, introducing non-causal decoding or future context—to examine the effects on performance?"

A direct comparison was not possible when we designed the experiments, because none of A²-LLM [2], FacePlex [3], or Moshi-Face [4] had public code. The A²-LLM code was released on November 10 [5], but without trained models or the dataset. As you suggest, we therefore approximated three of their core design choices on the same backbone, data, and cross-entropy (CE) predictor (Appendix I). Except for the last row, the rows of Table A differ only in the input or the motion codec and share the training setup. The last row is the full HiddenFace with the auxiliary losses.

Setting (CE only)DelayS-LVE↓S-MHD↓MOD↓Lip-W↓S-FDD↓L-FDD↓AVC↑MTM↓MSE↓FD↓P-FD↓rPCC↓
Agent audio codes09.2532.4264.1884.02124.7228.100.34767.90.19612.7513.710.142
Both speakers' codes09.3652.4444.1954.00725.0826.350.32969.30.19012.0112.970.127
Audio hidden only08.1942.2433.7063.61026.4028.530.34066.50.18712.1313.090.118
Audio + text hidden (ref.)08.1562.2463.7053.53325.4628.210.33967.40.18411.7912.750.116
ref. + token emb. (Moshi-Face-inspired)†08.0662.2243.6823.54425.3528.070.35664.30.18311.6612.620.115
— w/o last-row emb.08.1272.2363.6993.55025.4628.080.34367.80.18311.6912.650.116
ref. + 240 ms future240 ms7.8952.1993.6323.52924.4526.800.38857.50.18211.4112.380.111
Audio hidden + 240 ms future (FacePlex-inspired)240 ms7.8572.1923.6103.54024.8027.080.39556.20.18411.7312.700.114
Look-ahead codec, 1 token (Moshi-Face-inspired)160 ms8.0492.2353.6573.50323.8826.800.37656.20.18711.9212.900.115
Look-ahead codec, 2 tokens320 ms8.3162.2733.7013.62723.7026.940.39157.90.19912.7913.790.117
Look-ahead codec, 3 tokens480 ms8.4222.2913.8123.64023.7127.630.33566.40.19612.6013.590.116
HiddenFace (full objective)08.1892.2443.7113.45923.3226.520.38861.30.17910.9411.910.112

Table A. Comparisons on the same backbone, data, and CE predictor in the partner-prefix setting. Delay is the time by which a setting delays the output. † The audio code of the last row of a block is not yet reflected in the hidden states of that block and reflects audio up to 40 ms later than they do.

The first four rows of Table A change only the input of the predictor. Replacing the hidden states with the audio codes that Voila reads worsens S-LVE by 1.1–1.2 mm. This comparison reflects differences in available context as well as representation, because the hidden states also carry the agent's text and the preceding conversation.

AuthorsOpenReview ↗

Response to Reviewer PMQE (2/4): W1, Q1, Q3, continued

We start with future context. In FacePlex, each motion frame attends to about 240 ms of hidden states on either side, and speech and motion are both emitted 240 ms later ([3], Sec. 3.1–3.3). We approximated this by letting the predictor also read the next 240 ms of hidden states and delaying the output by 240 ms. S-LVE then improves by 0.33 mm with audio hidden states only, as in FacePlex, and by 0.26 mm with our input.

Next is non-causal decoding. Among the three methods, the non-causal component is the face codec of Moshi-Face ([4], Sec. 4). Its look-ahead is not reported. We therefore gave the decoder of our motion codec a look-ahead of one, two, and three tokens (a delay of 160, 320, and 480 ms) and retrained the codec and the predictor. With one token, S-LVE improves over the reference by 0.105 mm and MTM also improves, but MSE, FD, and P-FD worsen. With two or three tokens, S-LVE is worse than the reference.

Both choices delay the face by 160–240 ms. Because the speech must be delayed equally to stay in sync with the face, this delay is added to the agent's whole response. In Table C (Q5), HiddenFace emits the audio and face of a block a median of 9.4 ms after the block ends. Adding the look-ahead of 160–240 ms to this measured median gives an estimated 170–250 ms.

Another core choice of Moshi-Face is the input of its face module, which adds the embeddings of the text and audio tokens generated at the same step to the hidden state ([4], Eq. 2). We approximated this by adding the embedding of the agent's audio code generated in each row to the hidden state of that row (row "ref. + token emb. (Moshi-Face-inspired)"). S-LVE then improves by 0.09 mm without any added waiting. The hidden states of a block have not yet read the code of its last row. When only the embeddings of the codes they have already read are added (row "w/o last-row emb."), S-LVE is not distinguishable from the reference. Giving the last-row code as well would require redesigning the input for the token layout of each backbone, so we keep the default model on hidden states only.

In sum, the three choices we approximated improve S-LVE by 0.09–0.33 mm. Except for the 0.09 mm obtained without delay, these gains require delaying the output by 160–240 ms, whereas the default model needs neither generated-token inputs nor additional look-ahead. We therefore judged that they do not justify delaying the whole response by that amount.

Finally, we address joint fine-tuning of the backbone, the core choice of A²-LLM (Q1). We have not tried jointly fine-tuning Voila. The public Voila repository [6] provides inference code only, so training it jointly while preserving its speech ability would require rebuilding its original text and audio training losses and data.

Freezing the backbone has benefits in both computation and deployment. We extract and store the hidden states once and train only the 67.9M-parameter predictor, so training needs no forward or backward pass through the backbone. The speech and dialogue abilities of the deployed LALM also remain unchanged by construction.

We have no direct measurement of the performance difference. The two methods that fine-tune jointly, however, give some indication of the condition you state. A²-LLM reports catastrophic forgetting in joint training and uses a three-stage curriculum to prevent it ([2], Sec. 5.3). In Moshi-Face, speech naturalness dropped after fine-tuning (UTMOS from 3.08 for the original model to 1.75, [4], Table 2). On the performance side, both papers report that the variant with a frozen backbone synchronizes worse than the fine-tuned model ([2], Table 6: Peak Align 515 vs. 114 ms; [4], Table 2: LSE-D 11.8 vs. 8.76 in teacher-forced evaluation).

Our results show that the lip accuracy of Table 1 can be reached from hidden states alone with the backbone frozen and its speech generation unchanged. Whether joint fine-tuning can go beyond this remains future work.

AuthorsOpenReview ↗

Response to Reviewer PMQE (3/4): W2, Q2

W2 (semantic appropriateness of the reactions)
"The Voila backbone should provide access to semantic information, but the paper does not directly evaluate whether the generated reactions are semantically appropriate."

The paper indeed does not directly evaluate whether listening responses are semantically appropriate, and Sec. 5 states this as a limitation. We agree, however, that the expectation that Voila provides semantic information is close to the motivation of this work.

We therefore tested the step before semantic appropriateness: whether the listening face depends on what the partner actually said. HiddenFace does not receive the partner's audio, so its listening motion can depend on the partner's speech only through the hidden states computed by the frozen Voila. In each test conversation, we replaced only the partner's audio with that of a different speaker from another conversation whose timing of speech and silence is most similar (median timing correlation 0.76), and recomputed the hidden states. We then compared the two outputs with the motion of the real listener in frames where the agent listens and both partners speak, so that the speaking and listening states are the same in the compared frames.

Replacing the partner increases the expression-coefficient MSE by 6.5% (95% confidence interval [5.4, 7.6], Appendix J). In frames where the agent speaks and both partners are silent, in contrast, the difference in expression is not distinguishable from zero (0.4%). The partner's speech is thus reflected more in the listening face than in the speaking face. This test does not separate the content, prosody, and voice of the partner's speech, so we do not claim that it assesses semantic appropriateness.

Q2 (determinism and sampling)
"Is HiddenFace's main facial-motion branch deterministic? The paper appears to use greedy decoding for token selection. How would introducing randomness through sampling affect the results?"

Yes. The predictor selects the code of each residual depth greedily and is therefore deterministic for a given sequence of hidden states (Sec. 3.3). Rerunning the model in a new process reproduced the test outputs bit for bit. The paper did not test sampling. Following your question, we kept the checkpoint and inputs of Table 1 (audio encoded once). The only change was that the code of every depth was drawn from a softmax with temperature τ\tau (Table B, Appendix J). Each temperature is a single sampling run.

DecodingS-LVE↓S-MHD↓MOD↓Lip-W↓S-FDD↓L-FDD↓AVC↑MTM↓MSE↓FD↓P-FD↓rPCC↓
UniLS (released), 160 ms streaming (Table 1)8.6042.4843.7253.94024.5822.990.30558.40.24115.3216.440.135
HiddenFace, τ=1.0\tau = 1.08.4022.3543.9233.57020.9419.460.27178.80.1738.479.480.105
HiddenFace, τ=0.7\tau = 0.78.1492.2843.7743.46820.9020.510.32470.00.1658.529.490.107
HiddenFace, τ=0.5\tau = 0.58.0722.2493.7173.42821.3622.400.35466.50.1648.949.890.109
HiddenFace, greedy (Table 1)8.2292.2473.7303.45223.3326.880.37463.90.17811.0011.960.113

Table B. Sampling at temperature τ\tau in place of greedy decoding, with the same checkpoint and inputs.

With sampling, the metrics that measure the distribution gap to human motion (S-FDD, L-FDD, FD, P-FD) improve at all three temperatures, whereas AVC and MTM, which measure synchrony with speech, worsen as the temperature rises. Lip accuracy is comparable to greedy decoding at temperatures 0.5 and 0.7 and worse at 1.0. Sampling thus trades synchrony for distribution match, and the result is sensitive to the temperature, which has to be tuned with care.

The comparison with the released streaming UniLS does not rely on greedy decoding, because HiddenFace remains better on eight metrics, including S-LVE, at all three temperatures. We keep greedy decoding in Table 1 because it gives the same motion for the same conversation and the best synchrony. Greedy decoding also has no hyperparameter to tune, so the results of Table 1 do not depend on how a temperature is chosen.

Reviewer_JevdOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents HiddenFace, a method for generating conversational facial motion directly from the hidden states of a frozen large audio-language model (Voila). The model predicts facial motion in 160 ms blocks using audio and text hidden states together with previously generated motion tokens. A causal motion tokenizer and auxiliary training losses are used to improve motion quality. The main goal is to support streaming facial animation during both speaking and listening without reprocessing generated speech or introducing additional look-ahead.

Critical strengths
  • The most appealing aspect of HiddenFace is its simple interface with a frozen LALM. Generating facial motion from hidden states already available during speech generation avoids the need for a separate speech encoder or backbone modification. The block-causal design and fixed-size motion memory also make the approach attractive for real-time applications.
  • The evaluation is reasonably comprehensive, covering both speaking and listening behavior on 563 held-out conversations. Under the 160 ms streaming setting, HiddenFace improves several metrics over UniLS, including AVC (0.374 vs. 0.305), and achieves better dialogue-level distribution metrics.
Critical weaknesses
  • The contribution relative to recent LALM-based motion generation work could be better established. A2−LLMA^2-LLM (Hu et al., 2026) already uses LALM hidden states and motion history for facial motion prediction. FacePlex (Lim et al., 2026) and DyaPlex (Nagano et al., 2026) also explore related speech-motion generation settings, although their architectures and target modalities differ. HiddenFace makes useful design choices by keeping the backbone frozen, relying only on hidden states, and avoiding look-ahead. However, it is not entirely clear how much of the contribution comes from these particular constraints rather than the underlying prediction architecture. A more direct comparison or a clearer analysis of these differences would help establish the novelty.

  • The baseline comparisons do not fully isolate the benefit of hidden-state conditioning. In Table 1, several methods originally designed for offline inference are evaluated using restricted 160 ms audio prefixes without causal retraining, while DualTalk additionally loses its partner-motion input. Furthermore, HiddenFace's reported 5.8 ms latency uses CUDA graphs, whereas the baselines use their released implementations without equivalent optimization, making the magnitude of the efficiency improvement difficult to interpret.

  • The evidence for natural listening behavior is less convincing than the speaking results. The human study uses 30 clips selected based on content and audio from self-play conversations. HiddenFace receives the highest mean ratings, but its improvements over DualTalk are not statistically significant for lip synchronization, listening, or overall naturalness. UniLS also performs better on streaming L-FDD (22.99 vs. 26.88), suggesting that HiddenFace does not consistently improve listening motion across metrics. More importantly, the current evaluation does not directly test whether the generated listening expressions are appropriate for the conversational content. This seems particularly relevant given the motivation for using LALM hidden states.

Questions for authors
  • How would the main quantitative results change if the Voila hidden states were extracted using strictly causal, prefix-only audio tokenization? Could the end-to-end streaming analysis also include the intervals excluded from the steady-state measurements?
  • Could the authors clarify how much of HiddenFace's advantage comes from the hidden-state representation itself versus the additional conversational information it contains? A more controlled comparison with an LALM-based method or a causally adapted baseline would also be helpful.
  • How sensitive are the human-study results to the selection of clips? Would the conclusions hold for randomly sampled conversations, and is there a way to evaluate the semantic appropriateness of listening expressions rather than perceptual naturalness alone?
8 comments
AuthorsOpenReview ↗

Response to Reviewer Jevd (1/6): Q1

Thank you for the time and care you put into this review. We have uploaded a revised paper with the changes marked in blue. The sections, tables, and appendices cited below refer to this revision, and Appendices H, I, and J are new.

Q1 (prefix-only tokenization and the intervals outside the steady state)
"How would the main quantitative results change if the Voila hidden states were extracted using strictly causal, prefix-only audio tokenization? Could the end-to-end streaming analysis also include the intervals excluded from the steady-state measurements?"

In short, we repeated training and evaluation with prefix-only tokenization. When both speakers are encoded from prefixes, which is the strictly causal setting you ask about, the model retrained in this setting remains better than streaming UniLS on eight of the 12 metrics but is less accurate in S-LVE and MOD. When only the partner is encoded from prefixes, which is how the partner's audio is encoded in real-time operation, the comparison with each baseline of Table 1 gives the same result as before on all 12 metrics.

In Table 1, the audio of each speaker in the real conversations was encoded once per recording with the released Voila [1] tokenizer ("encoded once" in Table A). HiddenFace is causal at the token level (Appendix C). This tokenizer, however, is not a streaming codec, and in our measurement a code frame depends on up to 160 ms of later audio.

We therefore re-extracted the Voila states of both the training and test sets with prefix-only tokenization, in which each code frame is encoded from the waveform up to the end of that frame ("prefix-only" in Table A). We then retrained the model with the same setup.

We use two settings because whether this look-ahead departs from real-time operation depends on the speaker. In real-time operation, the partner's audio is encoded from only the audio received so far. The agent's codes are instead generated by Voila and not encoded from a waveform, so they involve no tokenizer look-ahead. In the partner-prefix setting, we therefore encode only the partner from prefixes and keep the agent's codes encoded once as the proxy for the generated codes. In the both-prefix setting, both speakers are encoded from prefixes, which is the strictly causal condition in which every input reflects only the audio up to the end of the current block.

Tokenization (partner / agent)ModelS-LVE↓S-MHD↓MOD↓Lip-W↓S-FDD↓L-FDD↓AVC↑MTM↓MSE↓FD↓P-FD↓rPCC↓
UniLS [2] (released), 160 ms streaming (Table 1)–8.6042.4843.7253.94024.5822.990.30558.40.24115.3216.440.135
Prefix-only / prefix-onlyretrained8.7322.3383.9293.47423.2926.500.32774.60.18811.6712.660.116
Encoded once / encoded onceoriginal (Table 1)8.2292.2473.7303.45223.3326.880.37463.90.17811.0011.960.113
Prefix-only / encoded onceretrained (Table 1)8.1892.2443.7113.45923.3226.520.38861.30.17910.9411.910.112

Table A. Results under the three tokenization settings (partner / agent).

For each metric and each released baseline row of Table 1, we compute the difference between HiddenFace and the baseline and obtain its 95% confidence interval by a paired bootstrap over the continuous runs of the test split. The comparison counts as not different if this interval includes 0 and as better or worse otherwise.

In the both-prefix setting, four advantages of the retrained model over offline baselines are also no longer significant. The other comparisons give the same result as in Table 1.

The agent's codes in this setting, however, differ greatly from those Voila generates. When speech generated by Voila is re-encoded and compared with the generated codes, all four codebooks match in 24.5% of the speech frames if the audio is encoded once but in only 3.0% with prefix-only tokenization. We therefore regard this setting as a waveform-causal replay rather than as the condition of real-time operation.

Table 1 has new partner-prefix rows, and the settings and results are in Appendix H.

To include the intervals excluded from the steady state, we re-analyzed the live run of Appendix F over all blocks. We excluded only the first block, which contains Voila's initialization, and added the analysis of the remaining 999 blocks to Appendix F. The block-end-to-emission latency is the time from the end of a block's fourth row interval until both its audio and face are ready and the block is emitted. Its 50th, 95th, and 99th percentiles are 9.5, 881, and 1300 ms over all blocks and 9.4, 19.3, and 19.6 ms over the 796 steady-state blocks.

The 135 blocks emitted more than 160 ms after they ended all fall in the intervals in which Voila generates text tokens in succession (up to 1.6 s) and then processes the queued input. In these blocks Voila's own audio is late, and no block missed the 160 ms deadline from waiting for the face.

AuthorsOpenReview ↗

Response to Reviewer Jevd (2/6): W1 and first part of Q2

W1 and first part of Q2 (constraints, prediction architecture, and the role of the hidden states)
"HiddenFace makes useful design choices by keeping the backbone frozen, relying only on hidden states, and avoiding look-ahead. However, it is not entirely clear how much of the contribution comes from these particular constraints rather than the underlying prediction architecture."

HiddenFace contributes a predictor that works under these three constraints. We show that the accuracy of Table 1 can be reached without relying on backbone fine-tuning, additional inputs, or output delay, and we measure how much accuracy is given up by not using additional inputs and output delay.

Recent methods that generate faces from LALM representations train the backbone jointly with face generation (A²-LLM [3], Moshi-Face [4]), use generated tokens as additional input (Moshi-Face [4]), or delay the output to use later context (FacePlex [5]). As you say, the prediction architecture of A²-LLM is close to ours in outline, and the difference lies in the conditions under which it is used. A²-LLM generates the face while the model answers a question ([3], Sec. 3.3) and fine-tunes the LLM jointly with LoRA. Its authors report that synchronization collapses when the LLM is frozen and only the connector is trained, and conclude that fine-tuning is essential ([3], Table 6: Peak Align 515 ms, against 114 ms with fine-tuning).

DyaPlex keeps the speech backbone frozen as we do, but a motion tower of the same depth reads the hidden states of all 32 backbone layers at every layer and takes the partner's motion as input ([6], Secs. 3–4). Its quantitative evaluation is limited to body motion ([6], Sec. 5).

HiddenFace keeps the backbone frozen and reads only the hidden states already computed at the last layer. It receives no partner motion, does not wait beyond the current block, and generates the face during both speaking and listening.

A direct comparison was not possible. None of A²-LLM, FacePlex, DyaPlex, or Moshi-Face had public code when we designed this comparison, and the A²-LLM code released on November 10 [7] comes without trained models or the dataset. We therefore varied the input and the look-ahead one at a time on the same backbone, data, and cross-entropy (CE) predictor (Appendix I). Table B addresses Q2 through controlled ablations of selected design choices from LALM-based methods, rather than full reproductions of those methods. All rows are trained and evaluated in the partner-prefix setting explained under Q1. Except for the last row, which is the full HiddenFace with the auxiliary losses, the rows of Table B differ only in the input or the motion codec and share the training setup.

Setting (CE only)DelayS-LVE↓S-MHD↓MOD↓Lip-W↓S-FDD↓L-FDD↓AVC↑MTM↓MSE↓FD↓P-FD↓rPCC↓
Agent audio codes09.2532.4264.1884.02124.7228.100.34767.90.19612.7513.710.142
Both speakers' codes09.3652.4444.1954.00725.0826.350.32969.30.19012.0112.970.127
Audio hidden only08.1942.2433.7063.61026.4028.530.34066.50.18712.1313.090.118
Audio + text hidden (ref.)08.1562.2463.7053.53325.4628.210.33967.40.18411.7912.750.116
ref. + token emb. (Moshi-Face-inspired)†08.0662.2243.6823.54425.3528.070.35664.30.18311.6612.620.115
— w/o last-row emb.08.1272.2363.6993.55025.4628.080.34367.80.18311.6912.650.116
ref. + 240 ms future240 ms7.8952.1993.6323.52924.4526.800.38857.50.18211.4112.380.111
Audio hidden + 240 ms future (FacePlex-inspired)240 ms7.8572.1923.6103.54024.8027.080.39556.20.18411.7312.700.114
Look-ahead codec, 1 token (Moshi-Face-inspired)160 ms8.0492.2353.6573.50323.8826.800.37656.20.18711.9212.900.115
Look-ahead codec, 2 tokens320 ms8.3162.2733.7013.62723.7026.940.39157.90.19912.7913.790.117
Look-ahead codec, 3 tokens480 ms8.4222.2913.8123.64023.7127.630.33566.40.19612.6013.590.116
HiddenFace (full objective)08.1892.2443.7113.45923.3226.520.38861.30.17910.9411.910.112

Table B. Comparisons on the same backbone, data, and CE predictor in the partner-prefix setting. Delay is the time by which a setting delays the output. † The audio code of the last row of a block is not yet reflected in the hidden states of that block and reflects audio up to 40 ms later than they do.

AuthorsOpenReview ↗

Response to Reviewer Jevd (3/6): W1 and first part of Q2, continued

The first four rows of Table B address the first part of Q2: how much of the advantage comes from the hidden-state representation itself and how much from the conversational information it contains. Adding the partner's audio codes to the agent's does not improve S-LVE, whereas giving the same audio of both speakers as Voila's audio hidden states improves it by 1.18 mm (95% confidence interval [1.10, 1.25]). The gain in S-LVE thus cannot be explained by the partner's codes alone and appears only when the hidden states are used. These states, however, also reflect the agent's text and the preceding conversation, so their information is not identical to that of the codes.

The dialogue-level distribution metrics behave differently. FD and P-FD improve once the partner's codes are added and do not change further when these codes are replaced by hidden states. L-FDD during listening is better with both speakers' codes than with the audio hidden states (26.35 vs. 28.53). With the auxiliary losses, the full predictor reaches 26.52 on hidden input, which is similar. We discuss the effect of decoding under W3.

Conversely, with the input fixed and only the training objective changed, the CE-only predictor and the full predictor with the auxiliary losses do not differ in S-LVE (8.156 vs. 8.189). With the predictor architecture fixed, the hidden-state input gives more accurate lip positions (S-LVE) than the tested audio-code inputs, and the auxiliary losses improve S-FDD, FD, AVC, and MTM without a distinguishable change in S-LVE. The memory of the predictor is ablated separately in Table 4.

Finally, we examined how much each of the three constraints limits the accuracy. We relaxed the hidden-states-only constraint by adding the embedding of the agent's audio code generated in each row to the hidden state of that row, as the face module of Moshi-Face does ([4], Eq. 2). S-LVE then improves by 0.09 mm without any added waiting. The hidden states of a block have not yet read the code of its last row. When only the embeddings of the codes they have already read are added (row "w/o last-row emb."), S-LVE is not distinguishable from the reference. Giving the last-row code as well would require redesigning the input for the token layout of each backbone, so we keep the default model on hidden states only and report this result in Appendix I.

We relaxed the no-look-ahead constraint in two ways. Reading the next 240 ms of hidden states and delaying the output by 240 ms, as in FacePlex [5], improves S-LVE by 0.26–0.33 mm. Giving the motion decoder a look-ahead of one token (160 ms), following the non-causal face codec of Moshi-Face ([4], Sec. 4), improves it by 0.105 mm, and with two or three tokens S-LVE is worse than the reference. Both choices delay the face by 160–240 ms. Because the speech must be delayed equally to stay in sync with the face, the agent's whole response is delayed.

We did not relax the frozen-backbone constraint ourselves, because the public Voila repository [8] provides inference code only. The most relevant published evidence is Moshi-Face ([4], Table 2). There, joint fine-tuning improved lip sync (LSE-D 8.76 vs. 11.8 in teacher-forced evaluation) but lowered speech naturalness (UTMOS 1.75, against 3.08 for the original model).

In sum, relaxing the two constraints we tested (hidden states only, no look-ahead) improves S-LVE by at most 0.33 mm.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response: Summary of Our Responses and the Revision

We thank the three reviewers and the Area Chair.

Our responses are long, so we summarize the main points by topic. Labels in parentheses point to the detailed answers, and changes in the revision are in blue.

HiddenFace is meant to add a face to a real-time voice agent, using only the hidden states already computed by a frozen LALM and without waiting for future speech.

1. Waveform causality of the Table 1 inputs (X1Ah W1, Q1, Q2 / Jevd Q1)

The states in Table 1 were extracted from codes obtained by encoding each recording once, and in our measurement they can reflect up to 120 ms of audio beyond the block. HiddenFace is causal at the token level and this dependence comes from the tokenizer (Appendix C of the submission). We therefore repeated training and evaluation with prefix-only tokenization, so that the inputs are also causal at the waveform level.

When only the partner's audio is encoded from prefixes (partner-prefix), the comparison with each released baseline gives the same result as before on all 12 metrics. When both speakers are encoded from prefixes (both-prefix), the retrained model remains better than streaming UniLS on eight metrics but is less accurate in S-LVE and MOD. UniLS also remains better in L-FDD and MTM, as in Table 1.

This change comes from the agent's own codes, which are the only difference between the two settings. In real-time operation, however, these codes are generated by Voila and not encoded. For speech generated by Voila, codes encoded once match the generated codes in 24.5% of the speech frames, against only 3.0% for prefix-only codes. By this measure, the partner-prefix setting is the closer of the two to real-time operation, and we therefore added it to Table 1. HiddenFace also received higher mean ratings than baselines run offline on the full audio in the self-play human study, where the agent's codes are actually generated.

2. Latency of the whole system (X1Ah W2 / Jevd Q1 / PMQE Q5)

We now report every block of the live run, including the startup and the text-generation stalls. Of the 999 blocks, 135 (13.5%) were emitted more than 160 ms after they ended. All of them fall in intervals in which Voila generates text tokens in succession, and no block missed the 160 ms deadline from waiting for the face.

After posting our responses, we also ran the system without the facial branch (three runs each). There are 111–132 late blocks without it and 115–134 with it, and the two ranges overlap. In the remaining blocks, the facial branch adds 6–14 ms at the 95th and 99th percentiles.

3. Comparison with related methods (Jevd W1 / X1Ah W2 / PMQE W1)

Recent methods train the backbone jointly, use generated tokens as additional input, or delay the output to use later context. HiddenFace is designed to work without these three, keeping the LALM frozen and not waiting for future context.

None of these methods had public code when we designed the experiments. We therefore compared their design choices one at a time on the same backbone, data, and predictor instead of reproducing the full methods. Replacing the hidden states with audio codes worsens S-LVE by 1.1–1.2 mm, whereas adding the choices of the other methods improves it by 0.09–0.33 mm. Except for the 0.09 mm obtained without delay, these gains require delaying the whole response by 160–240 ms. We judged that they do not justify this delay in a real-time conversation.

We keep the backbone frozen to preserve its speech ability. Methods that fine-tune it jointly report a loss of this ability, and we could not try it ourselves because Voila provides inference code only.

In sum, HiddenFace contributes a predictor that reaches the accuracy of Table 1 under these three conditions, which let a face be added without fine-tuning the LALM or waiting for future context.

4. Baseline conditions (Jevd W2 / X1Ah W2)

After UniLS is fine-tuned on streaming input, the original and partner-prefix HiddenFace both remain better on nine metrics including S-LVE. UniLS is better in MOD, L-FDD, and MTM. The ground-truth partner motion does not improve the lip accuracy of DualTalk. Without CUDA graphs, HiddenFace's step time (23.0 ms) is still shorter than that of every streaming baseline.

5. Evidence for listening (Jevd W3, Q3 / PMQE W2, Q2)

HiddenFace again has the highest mean score on randomly sampled clips. With sampling, L-FDD improves to the level of streaming UniLS at the cost of synchrony with speech. We do not claim semantic appropriateness, but the listening face does depend on the partner's speech. Replacing the partner's audio with that of a different speaker increases the expression error in listening frames by 6.5%.

6. Ethics (Jevd)

Because the participants' consent statement does not address the synthesis of new speech specifically, we restricted their voices to non-commercial research and revised the ethics statement accordingly.

We are happy to answer any further questions.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.