acceptodds
Under review as a conference paper at ICLR 2027

TiCo: Time-Controllable Spoken Dialogue Model

Abstract

Existing spoken dialogue models (SDMs) can generate high-quality responses, yet they struggle to control response duration. Such control is important for managing service costs, improving user experience, and supporting applications with duration requirements. We introduce TiCo-Bench, a benchmark for systematically evaluating duration control across four task categories and target durations of 10–60 seconds. Evaluations of open-source and commercial models reveal frequent failures to meet explicit duration constraints. To address this gap, we propose TiCo, a simple and efficient post-training method that incorporates duration estimation into semantic planning. TiCo uses Spoken Time Markers (STMs), such as <10.6 seconds>, to make estimated elapsed speaking time explicit during autoregressive generation, allowing the model to adjust the remaining content to meet the target duration. TiCo relies on self-generation and reinforcement learning with verifiable rewards, without requiring externally provided question-answer pairs. Across three backbone models, TiCo reduces duration error by 2.7–3.5, and the best TiCo model halves the error of the strongest baseline, while preserving response quality. Demo page: https://tico-9wu.pages.dev/

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
spoken dialogue models, time-controllable generation, reinforcement learning, speech language models

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_2EfVOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies explicit duration control in end-to-end spoken dialogue models and introduces TiCo-Bench to evaluate this capability across four task categories. The proposed method, TiCo, inserts Spoken Time Markers (STMs) into intermediate text to represent cumulative speaking time. A first-stage supervised fine-tuning process learns these markers from self-generated speech with automatic timestamps, followed by GRPO training that rewards duration compliance and discourages malformed markers and repetition. The speech generator remains frozen, and reinforcement-learning rewards are computed from text rather than synthesized audio. Experiments on three backbones report approximately 2.7–3.5× reductions in duration error, broadly maintained automatic quality scores, and generalization across duration targets and instruction forms.

Critical strengths
  1. Relevant problem and useful benchmark. The paper identifies a practical capability gap that existing SDMs handle poorly and provides a dedicated benchmark across multiple tasks. This establishes a useful evaluation resource beyond the proposed method.

  2. Efficient method with consistent gains. STM-based rewards avoid expensive audio synthesis during reinforcement learning. Substantial improvements across three backbones, supported by ablations and generalization experiments, provide credible evidence of the method’s effectiveness.

Critical weaknesses
  1. Insufficient validation of the STM-to-audio relationship. Rewards assess predicted timestamps rather than realized speech duration. Marker validity constraints do not ensure acoustic accuracy, and the reported marker–audio discrepancies warrant a clearer decomposition of planning error and realization error. Importantly, errors in generated markers should be distinguished from timestamp-label errors in the self-generated SFT data.

  2. Incomplete evidence for quality preservation under duration constraints. Aggregate GPT-score and UTMOS do not establish that duration-compliant responses remain complete and natural, especially for bounded instructions evaluated only with duration metrics. Reported degenerate endings are not adequately quantified, and the absence of human evaluation and confidence intervals weakens quality-preservation claims and close baseline comparisons.

Questions for authors
  1. Decomposition of duration error, with numbers rather than curves. The reward is defined against the final Spoken Time Marker, whereas the reported MAPE is computed against the realized audio duration. Figure 9 plots the error against the instructed duration and the error against the final STM, and the two curves nearly coincide, which suggests the planning error is small and the residual comes from the gap between planned and realized speech. However, this decomposition is only presented graphically. Could you report the two components numerically, namely the average error of the final STM relative to the instructed duration and the average error of the realized audio duration relative to the final STM? In addition, Table 8 reports a local alignment error of 2.65 s (Short) and 3.21 s (Long) aggregated over all markers. Could you report the same statistic for the final STM alone, since that is the quantity the reward actually verifies?

  2. Possible circularity between training labels and evaluation of marker accuracy. Stage 1 constructs supervision using word-level timestamps from the Whisper medium, and Appendix F.1 evaluates marker accuracy against ASR word-level timestamps. To what extent does the reported marker accuracy reflect true acoustic timing, rather than agreement between the model and Whisper's own timing estimates? Could you cross-validate with an independent aligner, for example, the Montreal Forced Aligner already used elsewhere in the paper for segment-level prosodic statistics?

  3. Quality metrics for the bounded instruction forms. Section 6.2 augments 100 audio queries with less than, more than, and in-between instructions and reports only MAPE and hinge MAPE, with hinge MAPE defined to be zero inside the admissible range. Could you additionally report the GPT score and UTMOS for these queries, in particular for the samples that fall inside the admissible range? Since hinge MAPE is identically zero there, the current metrics cannot establish whether an in-range response is complete and natural or degenerate.

  4. Empirical rate of the degenerate endings. The Limitations section notes that, in order to end on time, the model sometimes closes with a generic offer of further help or an unfinished clause, and that the automatic metrics do not explicitly penalize this. What is the empirical rate of this behavior on TiCo-Bench, and is it correlated with how far the requested duration lies from the model's natural response length?

  5. Statistical reliability of the main comparisons. Are the results in Table 1, Table 4, Table 7, and Table 12 obtained from a single run or averaged over multiple random seeds? If single-run, could you report confidence intervals or run-to-run variance, particularly for the comparison against the speaking-rate-controlled cascade, where the MAPE difference is about 1.5 points (16.19 vs. 17.67) while the UTMOS difference is 1.38 (4.04 vs. 2.66)? Note that the ±1 SEM shading in Figures 10 and 11 reflects sampling variability across evaluation samples and does not substitute for variance across training runs.

  6. Justification of the reward hyperparameters. The copy-penalty tolerance is set to τ = 0.5 s, which is considerably smaller than the reported marker alignment error of 2.65 s to 3.21 s. Could you explain how τ and the Gaussian tolerance σ were selected, and whether performance is sensitive to these values?

6 comments
AuthorsOpenReview ↗

Response to Reviewer 2EfV (1/3): W1, Q1 and Q2

We thank the reviewer for recognizing the practical value of TiCo and the relevance of the problem it addresses, as well as the consistent gains across three backbones, which the reviewer found to be credible evidence of the method's effectiveness. Below we address each concern.

W1: Validation of the STM-to-audio relationship

Thank you for raising this. W1 makes three points. Two are answered here; the decomposition into planning error and realization error is in Q1 below, and the independent aligner in Q2.

The reward is on the markers, not on the realized speech. We checked the realized audio at checkpoints along RL training.

Table R1-2EfV. Duration error and marker accuracy along RL training (retrained model, 320 prompts).

RL stepAudio vs. target (MAPE)Audio vs. final STM (MAPE)Pearson r (STM, audio)
0 (before RL, after Stage 1)44.0%11.9%0.89
20039.8%11.5%0.87
40017.6%9.7%0.95
60016.2%12.8%0.88
800 (after RL)15.4%12.3%0.94

RL reduces the error of the realized audio from 44.0% to 15.4%, while the gap between the marker and the audio stays at about 12%. The realization error is already present after Stage 1, and optimizing the markers does not widen it.

Label error against generated-marker error. We compared Whisper, which produces the labels, with the Montreal Forced Aligner (MFA) on the Stage 1 training data.

Table R2-2EfV. Whisper against MFA on the Stage 1 training data.

Boundary setCountMean abs. discrepancy
Word onsets (all)36,31274 ms
Word offsets (all)36,31268 ms
Offsets at marker positions4,99153 ms

At the positions that enter the supervision, the labels differ from MFA by 53 ms on average, below the 0.1 s resolution of the STM. The label error is thus roughly fifty times smaller than the error of the generated markers (Table R5-2EfV). The gap between marker and audio comes from the model's prediction of its own timing, not from noisy labels.

Q1: Decomposition of the duration error, with numbers

Planning error and realization error. For each reply we measure the realized audio against the instructed duration, the realized audio against the final STM (realization), and the final STM against the instructed duration (planning).

Table R3-2EfV. Decomposition of the duration error (Qwen2.5-Omni-7B + TiCo; TiCo-Bench, 2,000 prompts; MAE in seconds). The first two columns are the two curves of Figure 9, here over all four tasks.

Instructed durationAudio vs. instructedRealization: audio vs. final STMPlanning: final STM vs. instructed
10–20 s3.263.320.07
21–30 s3.954.090.22
31–40 s4.885.040.28
Beyond training range
41–50 s5.616.342.29
51–60 s6.336.427.53
Overall4.604.801.79

Up to 40 s, planning is nearly exact (the plan is within 1 s of the target in 97–98% of replies), so the error comes almost entirely from realization, as Figure 9 suggests. Beyond 40 s the two curves of Figure 9 still coincide, but the numbers show that planning is no longer exact: the final STM undershoots the target, and the audio running longer than the marker partly offsets it. Stage 2 samples targets only up to 40 s (Appendix C), so this range is outside training.

The final STM alone. Following Table 8, we report the error between the final STM and the ASR timestamp at the same position.

Table R4-2EfV. Alignment error of the final STM (MAE in seconds).

TaskShort (10–30 s)Long (30–60 s)
QA2.875.31
Reasoning3.245.92
Creative3.256.01
Summarization8.627.91
Overall3.595.82

The final marker has a larger absolute error than the all-marker average of Table 8, because timing error accumulates along the reply. In relative terms it is 13.8% of the elapsed time, consistent with the realization error above.

Q2: Cross-validation with an independent aligner

We repeated the analysis of Appendix F.1 with MFA in place of Whisper.

Table R5-2EfV. Marker error against two aligners (Qwen2.5-Omni-7B + TiCo, 320 replies; MAE in seconds).

MarkersCountSTM vs. WhisperSTM vs. MFAWhisper vs. MFA
All3,5582.492.520.08
Final only3104.074.100.07

The marker error is the same against either aligner, and the two aligners differ from each other by 0.08 s. The reported accuracy therefore reflects acoustic timing, not agreement with Whisper. Together with Table R2-2EfV, neither the training labels nor the evaluation depends on the choice of aligner.

AuthorsOpenReview ↗

Response to Reviewer 2EfV (2/3): W2, Q3 and Q4

W2: Evidence for quality preservation under duration constraints

Thank you for raising this. We agree that aggregate scores alone do not show that the replies stay complete, so we measured the endings directly on all 2,000 replies per model.

Table R6-2EfV. Ending types on TiCo-Bench (Qwen2.5-Omni-7B, 2,000 replies per model).

ModelUnfinished clauseGeneric offer of helpEither
Qwen2.5-Omni-7B (base model)0.0%91.0%91.0%
Qwen2.5-Omni-7B + TiCo13.2%36.0%38.5%

The generic offer is a habit of the base model, which ends 91% of its replies this way. TiCo reduces it to 36% and spends the time on the answer. The cost is the unfinished clause, which appears in 13.2% of TiCo's replies.

The unfinished ending can be largely removed. We retrained Qwen3-Omni with a system prompt that asks the model to end on a statement, without offering further help. The completion rate is the share of replies that end on a complete sentence.

Table R7-2EfV. Completion rate and duration error (TiCo-Bench, 2,000 prompts; MAPE of the first two rows from Table 1).

ModelCompletion rateMAPE ↓
Qwen2.5-Omni-7B + TiCo86.8%16.2%
Qwen3-Omni-30B + TiCo93.0%12.2%
Qwen3-Omni-30B + TiCo, trained to end on a statement99.2%17.7%

The completion rate rises from 93.0% to 99.2%, and the duration error (17.7%) remains far below the base model's 42.1%. The unfinished ending is therefore not inherent to duration control: it can be trained away, at some cost in duration accuracy. We are repeating this on Qwen2.5-Omni.

Confidence intervals and human evaluation. With a paired bootstrap over queries, the overall GPT-score does not differ between the backbone and TiCo (3.31 vs. 3.32; difference −0.01, 95% CI [−0.06, 0.04]). Intervals for the other comparisons are in Q5. A human evaluation is in progress, and we will add it to the revision.

Q3: Quality metrics for the bounded instruction forms

Section 6.2 tests whether duration control generalizes to instruction forms never seen in training, and TiCo reduces hinge MAPE by 2.4–3.4× on them. We agree that a zero hinge MAPE says nothing about the quality of an in-range reply. We are now running GPT-score and UTMOS on these replies, for each bounded form and separately for the in-range replies, together with the ending analysis of Tables R6-2EfV and R7-2EfV, and will report them in the revised version.

In the meantime, the demo page provided in the paper includes audio samples under less than, more than and in-between instructions, where the replies remain complete and natural.

Q4: Empirical rate of the degenerate endings

The empirical rate is in Table R6-2EfV above: 13.2% of TiCo's replies end on an unfinished clause and 36.0% on a generic offer, against 0.0% and 91.0% for the base model.

Relation to the natural response length. We take the base model's duration on the same query as the natural length.

Table R8-2EfV. TiCo's ending types by the distance between the requested duration and the natural length.

Requested minus natural lengthRepliesUnfinished clauseGeneric offer
More than 15 s shorter16910.1%14.8%
5 to 15 s shorter39512.9%20.5%
Within 5 s48818.4%37.9%
5 to 15 s longer45217.0%45.1%
More than 15 s longer4966.2%45.4%

The two endings behave differently. The generic offer depends on the direction: when the requested duration is shorter than the natural length, TiCo drops the offer (15–21%); when it is longer, there is room and the offer stays (45%). The unfinished clause does not grow with the distance: it is most frequent when the request is close to the natural length (17–18%) and least frequent at both extremes. Overall, the correlation between the absolute distance and a degenerate ending is −0.03.

AuthorsOpenReview ↗

Response to Reviewer 2EfV (3/3): Q5 and Q6

Q5: Statistical reliability of the main comparisons

All results in Tables 1, 4, 7 and 12 come from a single training run per model and a single generation pass, because RL post-training of 7B–30B SDMs is expensive. We add two measurements of uncertainty.

Sampling uncertainty and run-to-run variance. We report bootstrap 95% confidence intervals over the 2,000 queries. We also trained TiCo on Qwen2.5-Omni-7B a second time from scratch (new Stage 1 and Stage 2 runs).

Table R9-2EfV. Duration error with confidence intervals, and a second training run (TiCo-Bench, 2,000 prompts).

ModelMAPE ↓ [95% CI]
Qwen2.5-Omni-7B (base model)43.7% [42.0, 45.5]
Qwen2.5-Omni-7B + TiCo, second training run17.6% [16.6, 18.6]
Qwen2.5-Omni-7B + TiCo, Table 116.2%

The two training runs differ by 1.4 points, and the interval of one run is about ±1 point. The improvement over the base model is 26 points, more than ten times either source of variation. It is also replicated on three independently trained backbones with different architectures (Qwen2.5-Omni, Qwen3-Omni, Step-Audio-2-mini), each reducing MAPE by 2.7–3.5×.

Comparison with the speaking-rate-controlled cascade. We agree that the 1.5-point MAPE difference (16.19% vs. 17.67%) is within the run-to-run variation above, and we do not claim that TiCo is more accurate in duration than this cascade. The claim rests on naturalness.

Table R10-2EfV. Speaking-rate-controlled cascade and TiCo (from Tables 1 and 7).

SystemMAPE ↓UTMOS ↑ΔUTMOS
Cascade (Qwen), natural46.36%3.12–
Cascade (Qwen), adjusted17.67%2.66−0.46
Qwen2.5-Omni-7B43.33%4.09–
+ TiCo16.19%4.04−0.05

The absolute UTMOS (4.04 vs. 2.66) includes the contribution of each system's synthesizer: without any duration control the two systems already differ by 0.97. To isolate the cost of the control mechanism, we compare the change within each system. At comparable duration accuracy, rate adjustment lowers UTMOS by 0.46, whereas TiCo lowers it by 0.05. Both the 1.38 gap and the 0.46 drop are far larger than the sampling uncertainty of a mean over 2,000 utterances.

We will add confidence intervals to the tables and state the single-run setting in the revision.


Q6: Justification of the reward hyperparameters

τ and the alignment error measure different things. The copy penalty compares non-final markers with the instructed duration, not with the audio. It catches a textual shortcut: writing the target value early. τ therefore only needs to capture near-exact copies, and 0.5 s spans a few steps of the 0.1 s STM resolution. The 2.65–3.21 s alignment error is the distance between a marker and the audio, which the copy penalty does not use.

σ. σ = 5 s gives the main reward a useful slope across the 10–60 s range: the reward is 0.92 at a 2 s error, 0.61 at 5 s and 0.14 at 10 s.

Selection and sensitivity. Both are ordinary hyperparameters. We set them once and did not run a search, so we cannot report sensitivity. Table 12 does confirm the role of the copy penalty: the variants with it reach 15.38–19.07% MAPE, against 28.10–38.89% without it. We will state this in the revision.


Summary

We sincerely thank the reviewer for the time and patience in reading this long response, and for helping us make the manuscript more comprehensive. We will include these discussions (e.g. the decomposition of the duration error, the cross-validation with an independent aligner, the ending analysis and the confidence intervals) in the revised version. We are also running human evaluations.

Reviewer_XyW9OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary
  1. This paper investigates duration-controlled response generation in Spoken Dialogue Models (SDMs). The authors construct TiCo-Bench by combining queries from four existing datasets and adding duration constraints (10–60 seconds). They propose TiCo, a two-stage post-training framework: (1) SFT on self-generated responses annotated with Spoken Time Markers (STMs) obtained through ASR alignment, and (2) GRPO using rewards based on the difference between the final predicted STM and the target duration, without synthesizing speech during RL training.

  2. Experiments on three SDM backbones report 2.7×–3.5× reductions in duration error (MAPE) while maintaining comparable average response quality. The paper also provides comparisons with cascaded LLM-TTS systems and ablation studies.

Critical strengths
  • Training efficiency: By using text-based Spoken Time Markers (STMs) as a proxy for speech duration, TiCo enables reward computation without synthesizing audio during RL rollouts. This design reduces the computational overhead associated with speech generation during training, making the proposed approach more efficient and potentially scalable to larger SDMs.

  • Consistent empirical improvements: TiCo achieves substantial reductions in actual speech duration error (2.7×–3.5×) across three SDM backbones, while maintaining comparable average response quality. These results provide empirical evidence of the effectiveness of the proposed approach.

Critical weaknesses
  • The paper argues that speech duration cannot be reliably controlled through text length due to variations in phonetic composition, syllable structure, and prosody (Section 1). However, no controlled experiments quantify how significantly these factors affect actual duration. Comparisons between responses with identical word counts but different phonetic structures, as well as strong text-length-based baselines, are needed to establish the necessity of the proposed approach.

  • The idea of using intermediate markers to track generation progress is not new. Hansel [1], cited in Section 2.3, already employs tracking tokens for length control, yet no adapted baseline is provided. Moreover, VideoDubber [2] incorporates speech duration and remaining-time information into autoregressive generation but is not discussed. These omissions weaken the novelty justification and make it difficult to assess TiCo's methodological advances over prior work.

  • TiCo learns duration estimates from its own speech generator, but identical text can have different durations across TTS models, speakers, and speaking styles. Although the method is evaluated on three backbones, each is trained separately. The paper does not demonstrate whether its duration control generalizes across different speech synthesis conditions, despite this variability being a key motivation.

  • A major concern is that the model is optimized using rewards computed from timestamps generated by the model itself, rather than actual speech duration. This creates a potential circularity: the model may learn to generate timestamps closer to the target without necessarily improving the duration of the synthesized speech. Although the authors report average STM alignment errors of 2.65–3.21 seconds (Appendix F.1), these aggregate errors do not establish whether the final STM used for reward computation is sufficiently accurate, particularly for shorter duration targets. A more detailed analysis of final STM prediction errors, their distribution across duration ranges, and their evolution during RL training is needed to establish the reliability of this reward signal.

  • TiCo-Bench primarily combines queries from existing datasets with duration instructions, limiting its novelty as a benchmark.

Questions for authors
  1. What is the fundamental methodological difference between TiCo and prior approaches such as Hansel[1] and VideoDubber[2]? Why are adapted baselines from these methods missing?

  2. What prevents the model from optimizing its self-generated STMs rather than actual speech duration? Can you demonstrate that STM accuracy does not deteriorate during RL training?

  3. Since GRPO relies on STMs predicted by the same model being optimized, how reliable are these predictions as reward signals? In particular, how often would STM-based rewards rank candidate responses differently from rewards computed using actual synthesized audio durations? Can you quantify this discrepancy and investigate whether it increases during RL training?

References:

[1] Song, S., Lee, J. and Ko, H. (2024) ‘Hansel: Output Length Controlling Framework for Large Language Models’, arXiv preprint, arXiv:2412.14033

[2] Wu, Y., Guo, J., Tan, X., Zhang, C., Li, B., Song, R., He, L., Zhao, S., Menezes, A. and Bian, J. (2023) ‘VideoDubber: Machine Translation with Speech-Aware Length Control for Video Dubbing’, Proceedings of the AAAI Conference on Artificial Intelligence, 37(11), pp. 13772–13779.

5 comments
AuthorsOpenReview ↗

Response to Reviewer XyW9 (1/4): W1 and W2 (also Q1)

We thank the reviewer for recognizing the training efficiency of TiCo (Spoken Time Markers let the reward be computed without synthesizing speech during RL rollouts) and its consistent improvements across three backbones. Below we address each concern.

W1: Controlled evidence that text length does not determine duration

The paper motivates this in Sections 1 and 2.3. Table 13 also reports the speaking rate of each backbone, which varies across backbones (2.24 to 2.74 words per second on average) and within the same model (standard deviation 0.30 to 0.46). Here, we further ran controlled experiments to address it. Four of them are in our general response, and one more is below.

In the general response:

  • Table G1 shows that speaking style changes the rate: across 18 personas, Qwen3-Omni speaks at 2.07 to 3.28 words per second. End-to-end SDMs can respond in many different styles, which is motivation that TiCo estimates elapsed time instead of counting words.
  • Table G2 compares two predictors of audio duration under these styles: word count (17.1%, 16.7% error) and the Spoken Time Marker (9.1%, 13.7%).
  • Table G3 measures phonetic structure directly with forced alignment: a five-syllable word takes 3.7× as long as a one-syllable word.
  • Table G4 provides strong text-length baselines: a word budget in the prompt (33.1%) and a word-count tracker trained with our full pipeline (30.4%), against 17.6% with Spoken Time Markers.

Additional experiment: identical word counts, different duration. Among TiCo's 2,000 benchmark replies, replies with the same number of words differ widely in length.

Table R1-XyW9. Audio duration of replies with matched word counts (Qwen2.5-Omni-7B + TiCo). Each row pools replies within ±2 words of the stated count.

Words in the replyRepliesAudio duration, median5th–95th percentile
40 ± 210216.0 s12.9–20.5 s
80 ± 29931.3 s25.8–37.4 s
120 ± 29947.9 s40.9–58.5 s

Replies matched in word count still differ by 1.4–1.6× in duration. With different speaking styles the spread is larger.


W2 (also Q1): Relation to Hansel and VideoDubber

Thank you for pointing us to these two works. We agree that tracking progress with intermediate markers is not new in itself. TiCo differs from both in two ways: (1) it targets end-to-end spoken dialogue models, where the output is speech and not text; and (2) its markers carry elapsed speaking time, measured from the model's own audio. Estimating this time is not trivial, because it depends on how the words will be spoken. This challenge and the real elapsed time are the uniqueness of SDMs, not of text LLMs.

Hansel. Hansel periodically outputs hidden special tokens that track the remaining target length of the output, and fine-tunes the LLM with them. We adapted it to our setting and trained it with our full pipeline: each Spoken Time Marker is replaced by a word counter at the same position, and the model goes through the same SFT and RL with the same rewards.

Table R2-XyW9. Hansel-style word-count tracking against Spoken Time Markers (TiCo-Bench, 2,000 prompts per model).

ModelMAPE ↓Plan within 1 s of the target
Qwen2.5-Omni-7B (base model)43.7% [42.0, 45.5]–
Word-count tracker (reward on its counter)30.4% [29.2, 31.7]9%
Word-count tracker (reward on true word count)32.3% [30.9, 33.5]9%
TiCo (with STM; retrained)17.6% [16.6, 18.6]72%

The word-count tracker counts accurately, but the instruction is given in seconds, and the model does not learn to convert seconds into words: its planned length lands within 1 s of the target in 9% of replies, against 72% with seconds-valued markers.

As we argue in the general response, controlling length in the text alone is not the best fit for spoken dialogue models, because the same words can be spoken at different rates and in different styles (Tables G1–G4).

VideoDubber. VideoDubber controls translation length by conditioning an NMT decoder on duration-aware positional embeddings, using token durations estimated by a predictor integrated into the translation model. The translated text is then synthesized by a separate TTS model with additional duration adjustment. In contrast, TiCo directly controls the duration of speech generated by an end-to-end spoken dialogue model, learning elapsed time information from its own generated speech without a separate duration predictor. Moreover, while VideoDubber aims to match the duration of source utterances for video dubbing, TiCo follows explicit duration instructions to generate spoken responses of a specified length.

We thank the reviewer for bring up this, making our discussion more comprehensive. We will add both works to Section 2.3 with this discussion, together with other work related to VideoDubber, and add the Hansel-style baseline to the experiments.

AuthorsOpenReview ↗

Response to Reviewer XyW9 (2/4): W3

W3: Generalisation across voices and speaking styles

Thank you for raising this. We tested the two conditions that can vary within a single spoken dialogue model: the speaker and the speaking style. The speaking-style experiments (18 personas and 14 tone conditions) are described in our general response; here we report how well TiCo controls duration under them, together with a new experiment on an unseen voice.

A voice never used in training. Qwen2.5-Omni has two built-in voices, Chelsie and Ethan. TiCo's training data was generated with Chelsie only. At inference we switched to Ethan, who speaks 5% faster on identical text.

Table R3-XyW9. Duration error with the training voice and an unseen voice (320 prompts). Brackets are bootstrap 95% confidence intervals.

ModelMAPE ↓
Qwen2.5-Omni-7B (base model), Chelsie43.5% [39.3, 48.3]
Qwen2.5-Omni-7B (base model), Ethan41.6% [38.2, 45.4]
Qwen2.5-Omni-7B + TiCo, Chelsie (training voice)14.2% [12.6, 16.1]
Qwen2.5-Omni-7B + TiCo, Ethan (unseen voice)14.5% [13.0, 16.3]

The error rises slightly, by 0.3 points, which is within the confidence interval. The change is small because the two voices have similar speaking rates: the plan is made in the text and does not depend on the voice, so the audio shifts only by the ratio between the two rates.


Speaking styles. Style and persona are a harder test: on Qwen3-Omni the speaking rate changes a lot with them (Table G1).

Table R4-XyW9. Duration error under speaking-style changes (MAPE ↓). Each row compares the base model and TiCo on the same prompts.

SettingBase model+ TiCo
Qwen2.5-Omni, added tone instruction (Table G4)43.9%21.4%
Qwen3-Omni, 14 tone conditions (Setting B)42.9% [38.8, 47.1]13.9% [12.3, 15.5]
Qwen3-Omni, 18 personas (Setting A)35.3% [28.8, 43.0]12.8% [10.4, 15.4]
  • TiCo keeps its control in all three settings, with an error two to three times lower than the base model.
  • The Spoken Time Markers follow the change in rate (Table G2), which a fixed speaking rate cannot.

A different speech generator. In an end-to-end model the speech generator is part of the model, so replacing it means a different model. TiCo is a post-training method meant to be applied to each model: it needs no external annotation and learns timing from the model's own speech. This is why the three backbones are trained separately, and it is the intended use, not a limitation of the evaluation.

We will add the voice and style experiments to the revision.

AuthorsOpenReview ↗

Response to Reviewer XyW9 (3/4): W4 (also Q2 and Q3)

W4 (also Q2 and Q3): Reliability of the STM as a reward signal

Thank you for raising this important concern. We agree that a reward computed on the model's own timestamps could, in principle, be optimized without changing the speech. Our results indicate that this is not what happens: the duration error is measured on the realized speech, and TiCo reduces it by 2.7–3.5× across three backbones (Table 1). The markers are also grounded in real speech: Stage 1 learns them from word timestamps of the model's own generated speech.

Whether the Spoken Time Markers stay tied to the audio is central to TiCo, and the paper examines it in several places. The grounding term of the reward is designed to prevent reward hacking (Section 4.2, Appendix A), and its effect is ablated in Appendix H (Table 12). Appendix C.1 (Figure 9) compares the error against the instructed duration with the error against the final STM, and Appendix F.1 (Table 8) measures how well each marker aligns with the realized speech. Below we add three experiments on the points raised here: the accuracy of the final STM by duration range, its evolution during RL, and how an audio-based reward would rank the rollouts.


1. The final STM across duration ranges. The reward treats the final STM as a stand-in for the audio duration, so the relevant error is the distance between the two.

Table R5-XyW9. Error of the audio against the final STM, by requested duration (Qwen2.5-Omni-7B + TiCo; TiCo-Bench, 2,000 prompts).

Requested durationAudio vs. final STM (MAE)Audio vs. final STM (MAPE)
10–20 s3.32 s16.4%
21–30 s4.09 s13.4%
31–40 s5.04 s18.6%
41–50 s6.34 s13.0%
51–60 s6.42 s11.9%

The error in seconds is smallest for short targets and grows with duration. In relative terms it is 12–19% in every range, without a clear trend, so short targets are not a special weakness of the reward signal. This extends Figure 9 to the full benchmark with numbers, and restricts Table 8 to the final marker.


2. What prevents the model from optimizing its own STMs? (Q2, first part) Writing timestamps that climb to the target regardless of content is a real failure mode, and it is what the grounding terms of the reward remove (Appendix H, Table 12). With the main duration reward alone, the model exhibits reward hacking and reaches 32.2% MAPE. With all grounding terms, the error falls to 15.4%. On a larger evaluation set (1,440 prompts), the final STM of the main-only model lands 1.33 s from the target while its audio is 8.99 s from the STM; with all terms the audio is 4.91 s from the STM.


3. Does STM accuracy deteriorate during RL? (Q2) We evaluated checkpoints along RL training.

Table R6-XyW9. Duration error and marker accuracy along RL training (retrained model, 320 prompts).

RL stepAudio vs. target (MAPE)Audio vs. final STM (MAPE)Pearson r (STM, audio)
0 (before RL, after Stage 1)44.0%11.9%0.89
20039.8%11.5%0.87
40017.6%9.7%0.95
60016.2%12.8%0.88
800 (after RL)15.4%12.3%0.94

RL reduces the duration error from 44.0% to 15.4%, while the gap between the marker and the audio stays at about 12% throughout. The markers do not drift away from the audio as RL proceeds.


4. How often would an audio-based reward rank rollouts differently? (Q3) For 68 prompts we sampled 4 rollouts each, as in GRPO, at the same five checkpoints, and generated the speech for every one. The audio-based reward is the reward with the final STM in the duration term (Eq. (5)) replaced by the measured audio duration; everything else is unchanged. We then compared the ranking of the rollouts under the two rewards.

Table R7-XyW9. Agreement between the STM reward and an audio-based reward (retrained model, 68 prompts × 4 rollouts per step).

RL step0 (before RL)200400600800 (after RL)
Rollout pairs ranked the same way (chance 50%)81%74%68%58%64%
Audio-based reward (never optimized)0.300.380.630.670.64
Spread of the final STM within a group (sd)5.0 s4.3 s2.3 s1.4 s1.1 s

At the start of RL, where the learning signal is strongest, the two rewards order 81% of rollout pairs the same way. Agreement then falls, because the four rollouts of a prompt converge on the target (the spread of their final STM shrinks from 5.0 s to 1.1 s) and become near-ties that neither reward separates well. Meanwhile the audio-based reward, which training never sees, doubles from 0.30 to 0.64.

We will add Tables R5-XyW9, R6-XyW9 and R7-XyW9 to the revision.

Reviewer_LZCTOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces TiCo-Bench for evaluating spoken-response duration control across four task categories and targets of 10–60 seconds. Its proposed method, TiCo, learns Spoken Time Markers from self-generated speech and then applies reinforcement learning using marker-based duration rewards. Results across three spoken dialogue backbones show substantial reductions in duration error. However, the acoustic grounding of the reward, the independent contribution of the proposed mechanism, and the preservation of response quality remain insufficiently established.

Critical strengths

S1. Relevant problem with substantial reported improvements. Explicit duration control is useful for spoken interaction. TiCo reduces MAPE from 43.33% to 16.19% on Qwen2.5-Omni, from 42.06% to 12.18% on Qwen3-Omni, and from 65.57% to 22.40% on Step-Audio-2-mini, demonstrating consistent gains across different architectures.

S2. Practical training approach with useful diagnostic analyses. Self-generated timestamp supervision avoids externally authored responses, and marker-based rewards avoid waveform synthesis during RL. Analyses of marker alignment, text–speech consistency, reward components, and prosody provide useful evidence about the method’s behavior.

Critical weaknesses

W1. The “grounding” reward does not verify acoustic timing. Appendix A checks marker presence, monotonicity, repetition, and copying, but never compares the markers with realized speech during RL. Increasing timestamps ending at the requested duration can satisfy these checks independently of the content. Appendix F reports useful alignment results, but the 7.45-second local error on Short Summarization illustrates the remaining mismatch. Calibration before and after RL is needed to establish whether reward optimization preserves timing accuracy.

W2. Existing ablations do not adequately isolate the proposed mechanism. Figure 3 compares TiCo’s SFT–GRPO–CHORD pipeline against supervised instruction tuning, so its advantage cannot be attributed solely to STMs. Appendix G’s final-only variant also changes the applicability of intermediate-marker rewards. A word-count tracker trained with the same optimization procedure, together with a CHORD ablation, is needed to establish the specific contribution of speech-calibrated markers.

W3. Aggregate quality scores conceal task-level degradation. Table 5 shows that Qwen2.5-Omni’s Short QA score drops from 3.55 to 3.31 and Short Creative score from 3.59 to 3.40, despite an almost unchanged overall average. Appendix J also contains unfinished endings, which the limitations acknowledge are not explicitly penalized. Without uncertainty estimates, human evaluation, or completion-rate measurements, the claim of preserving response quality is insufficiently established.

W4. The strongest duration-controlled cascade is not evaluated. Appendix E.2 applies speaking-rate adjustment only to the Qwen-based cascade, although the GPT-based planner is substantially stronger in Table 1. Combining that planner with the same adjustment is necessary to assess the claimed advantage over cascaded systems. Naturalness comparisons across different synthesizers based only on UTMOS further limit the conclusion.

Questions for authors

See weaknesses

3 comments
AuthorsOpenReview ↗

Response to Reviewer LZCT (1/2): W1 and W2

We thank the reviewer for recognizing the practical value of explicit duration control and the consistent gains TiCo achieves across three architectures. We also appreciate the positive assessment of our training approach and diagnostic analyses.

We also posted a general response on why text length alone does not determine speech duration; its Table G4 extends the word-count comparison in W2 to prompts with an added tone instruction.

Below we address each weakness with new experiments.


W1: Does the grounding reward verify acoustic timing?

The paper ameasures marker accuracy against the realized speech (Appendix F.1, Table 8; Figure 9) and ablates each grounding term (Appendix H, Table 12). However, indeed, the grounding rewards are computed on the Spoken Time Markers alone. Avoiding synthesis inside the RL loop is a design choice and this leads to TiCo's training efficiency. This is discussed in Appendix A "sources of speech-rollout overhead". We also tried rewarding the real synthesised duration directly, and the reward did not improve during training. A reward on the markers is what made RL work in practice.

Below we provide the performance before, during and after RL. Because of the time limit, we sampled 320 prompts from the benchmark.

Table R1-LZCT. Duration error and marker accuracy along RL training (retrained model, 320 prompts).

RL stepAudio vs. Target (MAPE)Audio vs. Last STM (MAPE)Pearson r (STM, audio)
0 (before RL, after Stage 1)44.0%11.9%0.89
20039.8%11.5%0.87
40017.6%9.7%0.95
60016.2%12.8%0.88
800 (after RL)15.4%12.3%0.94

RL reduces the duration error from 44.0% to 15.4%, while the gap between the marker and the audio stays at about 12% throughout.

For Short Summarization, the 7.45 s error comes from the backbone model, not from the reward. The input is very long (98 s on average), as shown in Figure 7(b), and Qwen2.5-Omni-7B cannot handle it well. A stronger backbone does not have this problem: in Table 1, TiCo with Qwen3-Omni reaches 17.08% MAPE on Short Summarization, against 48.96% with Qwen2.5-Omni.


W2: Do the ablations isolate the proposed mechanism?

We agree that the final-only variant of Appendix G also removes the intermediate-marker rewards. However, the shuffled variant in the same appendix does not. It keeps the same response text and the same number of markers, and only shuffles the marker values, so they no longer match the actual speaking time measured from the model's own speech. It reaches 28.32% MAPE against 16.19% for TiCo (Table 10), so the gain comes from marker values that match the actual speaking time, not from the marker tokens themselves. Below we add the two experiments requested: (1) a word-count tracker trained with the same pipeline, and (2) TiCo without CHORD. All models use Qwen2.5-Omni-7B.

For the word-count tracker, we replace each Spoken Time Marker with a Word Count Tracker at the same position; for example, <4.8 seconds> becomes <14 words>. The reward converts the count to seconds with the corpus speaking rate (2.88 words/s). We retrain from scratch: word-count SFT, then word-count RL.

Table R2-LZCT. Word-count tracking against Spoken Time Markers, and the CHORD ablation (TiCo-Bench, 2,000 prompts per model). Brackets are bootstrap 95% confidence intervals.

ModelMAPE ↓Plan within 1 s of the target
Qwen2.5-Omni-7B (base model)43.7% [42.0, 45.5]–
Word-count Tracker (reward on its counter)30.4% [29.2, 31.7]9%
Word-count Tracker (reward on true word count)32.3% [30.9, 33.5]9%
STM without CHORD (retrained)17.9% [17.0, 19.0]84%
STM with CHORD (retrained)17.6% [16.6, 18.6]72%

Word count vs. seconds. Replacing seconds with words increases the error by 12.8 points [11.5, 14.1]. The word-count tracker counts accurately (its counter is within about 3 words of what it wrote). However, it must also learn to convert the requested seconds into a word count, and it does not: its planned length lands within 1 s of the target in 9% of replies, against 72–84% with seconds-valued markers, which need no conversion.

CHORD. Removing CHORD changes the duration error by only 0.3 points [−0.6, 1.2]. In our preliminary study, we found that CHORD helps stabilize training, but it does not play a critical role in duration control. This is why we describe it in the implementation details and not as part of the proposed method. How necessary CHORD is across different backbone models remains future work.

AuthorsOpenReview ↗

Response to Reviewer LZCT (2/2): W3, W4 and Summary

W3: Task-level quality, unfinished endings and uncertainty

Thank you for the careful reading. We will report the completion rate and uncertainty estimates in the revised version.

Uncertainty. We assess statistical uncertainty with a paired bootstrap: we resample queries, keeping each query's two scores together, and report the 95% confidence interval of the mean difference Δ (backbone minus TiCo). The overall GPT-score is statistically indistinguishable between the backbone and TiCo (3.31 vs. 3.32; Δ = −0.01, 95% CI [−0.06, 0.04], n = 2,000). Of the two task-level declines, only Short QA is statistically significant (Δ = 0.24, 95% CI [0.16, 0.34], n = 500), while the Short Creative difference is within the confidence interval (Δ = 0.19, 95% CI [−0.02, 0.40], n = 100).

Completion rate. We measured, over all 2,000 replies, how many end on a complete sentence.

Table R3-LZCT. Completion rate and duration error (TiCo-Bench, 2,000 prompts; MAPE of the first two rows from Table 1).

ModelCompletion rateMAPE ↓
Qwen2.5-Omni-7B + TiCo86.8%16.2%
Qwen3-Omni-30B + TiCo93.0%12.2%
Qwen3-Omni-30B + TiCo, trained to end on a statement99.2%17.7%

This is a trade-off we will report in the revised version. Under a time constraint the model has less room, and sometimes the clock runs out mid-sentence. This happens often for a closing remark ("If you want to know more about …"). In this case the main content is already generated before the cut, which is consistent with the overall GPT-score not dropping.

The cut-ending problem can be alleviated. When we retrain Qwen3 with a system prompt that explicitly asks the model to end on a statement, the completion rate rises from 93.0% to 99.2%, and the duration error (17.7% MAPE) remains far below the base model's 42.1%. We tried this on Qwen3-Omni first because it follows instructions better. We are now working on Qwen2.5-Omni.

We thank the reviewer for pointing this out. This finding leads to a better version of TiCo and a more complete discussion.


W4: The strongest duration-controlled cascade

We apply the same two-pass speaking-rate adjustment (Appendix E.2) to the GPT-based cascade.

Table R4-LZCT. Speaking-rate-controlled cascades (TiCo-Bench; the last two rows are from Table 1; all cascade rows use IndexTTS-2.5).

SystemMAPE ↓GPT-score ↑UTMOS ↑ΔUTMOS
Cascade (GPT), natural27.26%3.633.04–
Cascade (GPT), adjusted5.41%3.662.81−0.23
Cascade (Qwen), natural46.36%3.093.12–
Cascade (Qwen), adjusted17.67%3.072.66−0.46
Qwen2.5-Omni-7B (unmodified)43.33%3.314.09–
Qwen2.5-Omni-7B + TiCo16.19%3.324.04−0.05

The adjusted GPT cascade achieves the lowest MAPE, but requires synthesizing the entire response to measure its duration, then synthesizing it again at an adjusted rate. Adjusted playback therefore must wait for the complete first pass. TiCo instead plans duration during generation and requires no second synthesis pass.

Beyond the system-level UTMOS comparison, we report changes relative to each system's own baseline. Rate adjustment lowers UTMOS by 0.23–0.46, whereas TiCo lowers it by only 0.05. GPT-score changes are small across all methods (−0.02 to +0.03). Thus, the stronger cascade achieves tighter timing through two-pass correction, while TiCo substantially improves duration control in a single synthesis pass, with higher UTMOS and a smaller decrease relative to its backbone.


Summary

We thank the reviewer again for helping us make the manuscript more comprehensive. We will include these discussions (e.g. the word-count tracker, the ablations and the completion rate) in the revised version. We are still running experiments to complete the tables (e.g. Table R3-LZCT), and we are also running human evaluations.

Follow-up: human evaluation.

As promised, we ran a small-scale human evaluation to check whether the automatic quality metrics in the paper reflect human judgment. Three raters scored 100 responses for naturalness (MOS, 1–5) and content quality (1–5). Brackets are bootstrap 95% confidence intervals.

Correlation of the mean human score with the metric (Pearson / Spearman)Agreement between raters (mean pairwise Pearson)Reliability of the mean score, ICC(2,3)
Naturalness vs. UTMOS0.682 [0.51, 0.81] / 0.610 [0.45, 0.74]0.4360.605
Content quality vs. GPT-score0.675 [0.56, 0.77] / 0.647 [0.50, 0.76]0.6550.763

Both metrics correlate with the mean human score (Pearson 0.68), at a level comparable to the agreement between two human raters (0.44 for naturalness, 0.66 for content quality). This supports using UTMOS and GPT-score as the quality measures of the paper. We acknowledge that this is a small-scale human evaluation, and we will extend it in the revised version.

Comments on the paper

2 comments
AuthorsOpenReview ↗

General Response: why text length alone does not determine speech duration

We thank all reviewers for their comments. Two points underlie several of them, and we believe they are central to the paper. In this response, we also report experiments on text-length control.

1. TiCo is an efficient post-training framework for end-to-end SDMs. TiCo gives an existing end-to-end spoken dialogue model (SDM) control over its speaking duration. It needs no external annotation, learns timing from the model's own speech, and computes its RL reward on Spoken Time Markers without synthesizing speech (much slower, Appendix A). To our knowledge it is the first work on duration instructions for end-to-end SDMs. Cascaded systems are indeed strong, and they appear in our comparisons as reference points (we provide more cascaded baselines in the rebuttal). However, they fall outside the scope of end-to-end SDMs.

2. Text length alone does not determine speech duration. In end-to-end SDMs, the speech generator (i.e. the talker in Qwen models) is conditioned on more than discrete text (Eq. (2); in Qwen, on the thinker's hidden representations). The same words can therefore be spoken at different rates and in different styles. This flexibility is a key motivation for end-to-end SDMs over chaining an LLM with a TTS system, and it is why text length alone cannot reliably determine spoken duration.

The three experiments below quantify this.

Experiment 1. End-to-end SDMs speak in different styles and at different rates.

Setting A: 18 speaking personas. We give Qwen3-Omni a persona in the system prompt (bedtime storyteller, sports commentator, technical lecturer, …) and a text prompt asking it to talk about a topic: 144 responses per model.

Table G1. Lowest and highest persona means.

Range across 18 personasQwen3-Omni (base model)Qwen3-Omni + TiCo
Words per second2.17 – 3.322.07 – 3.28
Syllables per word1.22 – 2.101.25 – 2.40
Syllables per second3.31 – 4.533.40 – 4.94
Share of the audio spent in pauses16% – 35%14% – 26%

Style changes both the words the model chooses and how it delivers them.

Setting B. 39 TiCo-Bench prompts, each run under 14 tones: the plain prompt and 13 versions with one added instruction (speak slowly, sound excited, use technical terms, …), 546 replies in total.

Table G2. Error in predicting the audio duration of Qwen3-Omni + TiCo replies, from the word count (divided by the model's average speaking rate) and from the last time marker.

Predicted fromA: 18 speaking personasB: 14 tones
Word count ÷ average speaking rate17.1%16.7%
STM (the last time marker)9.1%13.7%

Experiment 2. In ordinary replies, duration depends on phonetic structure and delivery.

We force-aligned all TiCo benchmark replies with the Montreal Forced Aligner.

Table G3. Word duration by syllable count (Qwen2.5-Omni-7B + TiCo, TiCo-Bench).

Syllables in the wordWordsMean durationRelative to one syllable
1115,4560.23 s1.0×
234,8540.43 s1.9×
311,5920.58 s2.5×
43,2960.71 s3.1×
5+7720.86 s3.7×

A word count treats a five-syllable word and a one-syllable word as the same unit, although the first takes 3.7× as long to say. Across replies, the variation in seconds per word comes from the time spent on each syllable (63%), the number of syllables per word (27%) and pausing (11%).

Counting syllables instead of words does not solve this: the delivery still changes with tone and persona. In Experiment 1 the rate ranges from 3.31 to 4.94 syllables per second, and pauses take 14% to 35% of the audio (Table G1).

Experiment 3. In controlled experiment, word count falls short of time markers.

Table G4. Text-length control against Spoken Time Markers (Qwen2.5-Omni-7B; TiCo-Bench, 2,000 prompts; tone column: 48 prompts × 14 added tone instructions).

MethodMAPE ↓MAPE with an added tone condition ↓
Qwen2.5-Omni-7B (base model)43.7%43.9%
Target given as a word budget (no training; 320 prompts)33.1%–
TiCo (with Word-count marker, e.g. <14 words>; same pipeline)30.4%40.2%
TiCo (with STM; retrained; same pipeline)17.6%21.4%

The word budget converts the requested seconds into words and asks the base model for that many. The word-count tracker is trained with TiCo's full pipeline, with word counters in place of time markers. Both text-length methods stay above 30%, and with an added tone instruction the word-count tracker falls back to 40.2%, close to the base model.

Overall, these experiments show that (1) end-to-end SDMs speak in different styles and at different rates conditioned on the input, which is a motivation why researchers build end-to-end models and not cascaded systems, and (2) text-length control works for text LLMs (shown in literature), but is less suitable for controlling the duration of SDMs.

Dear reviewers,

We thank the reviewers for their positive assessment of our work: the relevance of explicit duration control for spoken dialogue models, the usefulness of TiCo-Bench as a dedicated benchmark, the training efficiency of TiCo, its consistent gains across three backbones with comparable response quality, and the diagnostic analyses, ablations and generalization experiments that support the method.

We sincerely appreciate the reviewers' help in improving the manuscript and its discussion. Many of the points raised are analyzed in the paper, as the reviewers also noted in their comments, and our replies point to the relevant sections. We have also run additional experiments to further support these claims, including but not limited to:

  • Word-count tracker vs. Spoken Time Marker (STM). We compared STMs with a Hansel-style word-count tracker trained with the same pipeline. The duration error was 30.4% MAPE with word counts and 17.6% with STMs. Further discussion of length control is in the general response.

  • Reliability of the STM. During RL, the duration error of the generated speech fell from 44.0% to 15.4%. The gap between the final STM and the actual speech duration stayed around 12%. The marker error was the same against Whisper and MFA.

  • Ablations and generalization. Removing CHORD changed MAPE by only 0.3 percentage points. TiCo also improved duration control for an unseen voice and across speaking tones and personas.

  • Response quality. We added confidence intervals and measured how replies end. Overall GPT-scores were statistically unchanged. Training Qwen3-Omni to end on a statement raised completion from 93.0% to 99.2%, with some loss in timing accuracy.

  • Baselines and prior work. We tested a stronger cascade with speaking-rate control, adapted Hansel to track word count, and discussed VideoDubber.

Our individual replies discuss these results and the limitations in detail. We would like to thank the reviewers again for their patience and effort to review our paper and response.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.