acceptodds
Under review as a conference paper at ICLR 2027

Laconic: Configurable low-rate speech representations

Abstract

Speech representations spend capacity along two axes: how many tokens are emitted per second, and how many dimensions each token carries. Yet they are usually characterized by token rate alone, with token dimensionality treated as a fixed architectural choice. We ask how capacity should be allocated between the two. We introduce Laconic, a continuous speech representation whose token rate is chosen at encoding time and whose token dimensionality is chosen at reading time, as a nested prefix of each token, so that a single encoder exposes a two-axis family of operating points. Mapping this space, we find that token rate and dimensionality are partially interchangeable between roughly 3 and 6 tokens per second, but the exchange reaches a temporal floor at 1 token per second, where no tested dimensionality recovers the lost content. Different information also occupies different regions of the space. Under our nested objective, linguistic content becomes readable from low-dimensional prefixes while speaker identity emerges only at high-dimensional ones, and is carried more compactly by a small set of utterance-level tokens. The same geometry governs generation. In zero-shot TTS, low-dimensional tokens are easiest to generate at higher token rates, high-dimensional tokens help as the token rate decreases, and speaker similarity and perceptual quality favor different dimensionalities. These results suggest treating speech representation capacity not as a single rate budget, but as an allocation across token rate and dimensionality.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
speech, representation learning, compression

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_h5yiOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

Laconic looks at a trade-off between token rate (ie, number of tokens transmitted per second) and token dimensionality, in the world of neural speech codecs. The paper trains models at fixed token rates and dimensionality, performing a grid search and discovering a region of consistently low error, with marked degredation as token rate gets extremely low (1 tok/s).

Critical strengths

This is potentially a very rich question, trading off temporal frequency for geometric structure. At low-enough rates, the representation has to encompass more and more temporal structure, and presumably would require higher dimensional vectors to do that well.

Critical weaknesses
  • There is extensive literature on the information content of speech for compression which should be relevant to the question at hand; some greater relationship drawn to the context beyond WavLM would be desirable. How much information is actually being carried in the tokens? What does that tell you about the allowed shape of the tokens? Are there known limits to information rate elsewhere in the speech coding literature which would be relevant to the question you are asking?

  • The motivation of the paper isn't clear. Why do the authors care whether token dimensionality is larger or smaller? Why do they care whether the token rate is higher or lower? My expectation is that the goal is to reduce context size for a downstream WavLM-like system; in which case, it would be helpful to see some end-to-end metrics that demonstrate the utility of swapping in the larger vectors at lower rate.

  • This work would benefit from a more information-theoretic discussion. Token rate and dimensionality both have implications for overall bitrate of the stream. There is extensive work on low-bitrate coding of speech signals, including noticing when a bitrate of a (sub)-vector can be reduced as low as zero (ie, dropping out entirely, and thus reducing the apparent dimension of the vector), for example in the Opus encoder. In this case, we can imagine that some lower-bound on token rate is needed to represent the signal, while the dimensionality of the representation varies dynamically as needed.

In the end, a single system is trained and tested; better theoretical grounding could help overcome that shortcoming. As it is, I'm left wondering whetehr these results are actually useful or an artifact of the particular system that was chosen for training.

Questions for authors
  • It would be helpful to better explain the purpose of this work. For example, it isn't clear to me that this is particularly helpful in the problem of low-rate speech coding (there is no discussion of overall bitrate), but seems more geared towards reducing the context-length of a downstream transformer by trying to reduce overall token count?

  • At the extreme, we can imagine sending a single second of audio by simply transposing it as a single 16k-dim'l vector... Does the system improve as we approach the dimensionality of the raw audio?

2 comments

Questions For Authors:

A-Q1: It would be helpful to better explain the purpose of this work. For example, it isn't clear to me that this is particularly helpful in the problem of low-rate speech coding (there is no discussion of overall bitrate), but seems more geared towards reducing the context-length of a downstream transformer by trying to reduce overall token count?

Yes, our work is geared towards reducing the context length of a downstream transformer by reducing overall token count. We will make this clearer in the introduction.

A-Q2: At the extreme, we can imagine sending a single second of audio by simply transposing it as a single 16k-dim'l vector... Does the system improve as we approach the dimensionality of the raw audio?

We indeed looked at representing the utterances at much higher dimensionalities, as shown in Figure 3 (a) row “WavLM-Stack”. For completeness, we also experimented with transposing all the features (whether WavLM or raw waveform samples) in a second to obtain 1 token per second at a dimensionality of 16k. That leads to much worse results: using the raw waveform at 16k gave 115.5% WER, and WavLM-Stack gave 11.7% WER, compared to Laconic WER which is below 4%.

Critical Weaknesses:

A-W1: There is extensive literature on the information content of speech for compression which should be relevant to the question at hand; some greater relationship drawn to the context beyond WavLM would be desirable. How much information is actually being carried in the tokens? What does that tell you about the allowed shape of the tokens? Are there known limits to information rate elsewhere in the speech coding literature which would be relevant to the question you are asking?

How much information is in the tokens and what is says about the allowed token shape

To quantify this, we use a single delta rule throughout: standardise each dimension, one absolute step Δ for each of the dimensions, rate = summed per-dimension empirical entropy, with the utterance tokens amortised. Measured: 5.95 bits/dim in the lower dimensions vs 4.24 in the higher dimensions.

Known limits elsewhere in the literature: We compare Laconic with other Codecs, namely Mimi and FlexiCodec in the table below. Uncoded bits: naive uncompressed representation. Entropy-coded: theoretical limit with ideal entropy coding. Gzip: actual bitrate obtained by compressing the representation with gzip.

Systemtok/suncoded b/sentropy-coded bits/sgzip b/sWERSIM-o
Laconic Q, d=643.132,1911,0891,1753.040.514
Mimi, 8 cb12.51,1001,0431,2494.180.570
FlexiCodec, 24 cb4.171,2131,1501,3223.840.573
Laconic Q, d=646.253,7912,0192,1682.550.531
Mimi, 32 cb12.54,4004,2335,1192.460.696

We add a bitrate table which highlights the information in tokens per second. At matched bitrate (~1,050 b/s) Laconic is the most intelligible of the three (WER 3.04 vs 4.18 and 3.84) and the least speaker-faithful (0.514 vs 0.570, 0.573).

A-W2: The motivation of the paper isn't clear. Why do the authors care whether token dimensionality is larger or smaller? Why do they care whether the token rate is higher or lower? My expectation is that the goal is to reduce context size for a downstream WavLM-like system; in which case, it would be helpful to see some end-to-end metrics that demonstrate the utility of swapping in the larger vectors at lower rate.

it would be helpful to see some end-to-end metrics that demonstrate the utility of swapping in the larger vectors at lower rate.

We indeed show this in the paper on both speech reconstruction (Figure 5), ASR (Figure 3b), and Text-To-Speech (Figure 6)

A-W3: This work would benefit from a more information-theoretic discussion. Token rate and dimensionality both have implications for overall bitrate of the stream. There is extensive work on low-bitrate coding of speech signals, including noticing when a bitrate of a (sub)-vector can be reduced as low as zero (ie, dropping out entirely, and thus reducing the apparent dimension of the vector), for example in the Opus encoder. In this case, we can imagine that some lower-bound on token rate is needed to represent the signal, while the dimensionality of the representation varies dynamically as needed.

This work would benefit from a more information-theoretic discussion.

We will include an information-theocratic discussion in-line with the results and table we include above, under the first weakness.

we can imagine that some lower-bound on token rate is needed to represent the signal, while the dimensionality of the representation varies dynamically as needed.

We believe this is exactly what we show in the paper. Understanding tasks such as ASR saturate with few dimensions (Figure 4). More demanding tasks such as reconstruction need more dimensions (Figure 5).

Reviewer_V3byOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This work proposes LACONIC, a continuous speech representation with configurable token rates and nested latent dimensions. Built on frozen WavLM features, it combines local tokens with utterance-level tokens and supports multiple configurations through shared speech decoders. The paper evaluates rate–dimension trade-offs through probing, reconstruction, TTS, and additional understanding tasks. The main contribution is a flexible representation architecture accompanied by extensive within-framework ablations. The experiments characterize LACONIC’s behavior, while the broader applicability and practical benefits of the resulting design recommendations remain insufficiently established.

Critical strengths
  1. It is interesting that this work proposed a tokenrate-dimension configurable speech representation framework.
  2. Ablation studies show the trade-off property of proposed representation, providing some insight for downstream configurations.
Critical weaknesses
  1. The incremental contribution and competitiveness against relevant continuous representations are insufficiently established.

    Several uncited works directly overlap with the paper’s methodological motivation: Semantic-VAE and HoliTok investigate semantic alignment and task-informed representation learning; WavCube and LoSATok study compact semantic–acoustic representations using feature reconstruction. These works provide relevant comparison points for reconstruction, understanding, and generative modeling. However, the paper neither clearly distinguishes its incremental contribution from these approaches nor directly compares representative LACONIC configurations with them under appropriately controlled conditions.

  2. The evidence does not sufficiently support the generality and reliability of the claimed trade-offs.

    The extensive ablations characterize LACONIC, whose supervision explicitly shapes information allocation. They do not establish that the principal findings generalize to independently designed SSL-derived continuous representations or VAEs. The existing external diagnostics are insufficient to support broad tokenrate–dimension conclusions or an intrinsic temporal floor. Although Appendix C.1 acknowledges feature dependence at 1 Hz, the stronger interpretation in section 4.3 ("the information itself is no longer preserved") is not established by the reported decoding failures. The practical benefit of joint configurability also remains unclear. Most downstream models use fixed configurations, including separately trained TTS generators. The dedicated-encoder controls are useful, but do not establish a substantial training or deployment advantage over fixed-configuration alternatives. Furthermore, averaging three waveform-rendering seeds does not assess variability across generator training runs, and uncertainty in the TTS WER differences is unreported. Independent verification and perceptual validation are particularly limited. The supplied manuscript provides no accessible links to implementation code, model weights, or audio demonstrations, and reports no subjective listening evaluation. Perceptual-quality conclusions rely on UTMOS predictions rather than listener judgments. Consequently, reviewers cannot readily reproduce the reported trends or inspect whether the metric differences correspond to audible improvements, degradation, or synthesis failures. Combined with the missing uncertainty analysis, these omissions substantially weaken support for the generation and perceptual-quality claims.

Questions for authors
  1. Could you please provide some train&infer codes and model weights for reproduction? or at least some listening demos for reconstruction and downstream TTS performance? Has any subjective listening evaluation been conducted?
  2. Appendix H reports several downstream benchmarks. Could the authors compare LACONIC with relevant continuous representations under a common evaluation protocol, accounting for differences in token rate, representation size, and downstream model capacity?
  3. For representative comparisons supporting the main TTS conclusions, do the trends persist across independently trained generators? Please report training-run variability and confidence intervals for the WER differences.
  4. What measurable benefit comes specifically from joint configurability, beyond the benefits of low-rate representations and reusing a tokenizer checkpoint, given that downstream models are trained separately for different configurations?
4 comments

Questions For Authors:

A-Q1. Could you please provide some train&infer codes and model weights for reproduction? or at least some listening demos for reconstruction and downstream TTS performance? Has any subjective listening evaluation been conducted?

Could you please provide some train&infer codes and model weights for reproduction?

We will publicly release everything needed for reproduction of our work (training & inference code and model weights) upon acceptance of the paper.

Some listening demos for reconstruction and downstream TTS performance?

The listening demo is here (https://laconic429.github.io/).

Has any subjective listening evaluation been conducted?

Yes we did a subjective listening evaluation with 9 participants, and will include it in the final manuscript.

Automatic metrics (UTMOSv1 and v2) and human listeners (CMOS) mostly agree. But they considerably disagree on Laconic flow decoder vs Laconic L2 decoder, as well as Sylber vs the ground truth.

Tested systemagainstCMOS ↑ΔUTMOSv1ΔUTMOSv2
FlexiCodec.ground truth+0.04 ± 0.32+0.06−0.05
Laconic, flow decoder.ground truth−0.37 ± 0.32−0.59−0.36
Sylberground truth−0.37 ± 0.39+0.04−0.03
Laconic, ℓ2 decoderground truth−1.06 ± 0.47−0.37−0.14
Laconic, flow decoderLaconic, ℓ2 decoder+0.57 ± 0.59−0.22−0.23
Laconic, flow decoderFlexiCodec, 4.17 tok/s−0.49 ± 0.41−0.65−0.31

Flexicodec is 4.17 tokens/sec (with 24 codebooks each), Sylber is 3.76 tokens/seconds (with 768 dimensions), and all Laconic models are 4.17 tokens/sec (with 512 dimensions).

A-Q2. Appendix H reports several downstream benchmarks. Could the authors compare LACONIC with relevant continuous representations under a common evaluation protocol, accounting for differences in token rate, representation size, and downstream model capacity?

We extend Table 21 of Appendix H which contains 4 standard downstream tasks from the common protocol of SiTok/DASB. For ASR, we observe that, at matched representation size, LACONIC is the best or tied for best, but at a much lower cost.

SystemTokens/s @ DimRepr. size (# values/s)ASR WER (%) ↓ASR number of training tokens (B)ASR compute (wall-clock training time in hours)ER accuracy (%) ↑Speaker Verification (%) ↓Keyword Spotting (%) ↑
LACONIC6.25 @ 5123,5453.73 ± 0.130.226.265.27.1197.3
LACONIC3.13 @ 5121,9454.66 ± 0.040.114.267.07.1597.4
Semantic-VAE40 @ 642,56018.97 ± 0.251.3927.849.66.1296.0
LoSATok25 @ 1283,2005.68 ± 0.080.8620.169.55.7798.3
HoliTok25 @ 1283,20018.88 ± 0.080.8718.657.26.7194.2
WavCube50 @ 1286,4003.75 ± 0.011.7339.365.73.6698.3

A-Q3. For representative comparisons supporting the main TTS conclusions, do the trends persist across independently trained generators? Please report training-run variability and confidence intervals for the WER differences.

do the trends persist across independently trained generators?

Yes. We retrained the generator from scratch with three training seeds at key cells (6.25 and 2 tok/s × d ∈ {64, 512}), giving 12 independently trained generators. We observe lower WER as we increase dimensions at high(6.25) tokens/sec. And this reverses at low(2) tokens/sec.

setrate (tokens / s)Δ(512−64) seed 0Δ(512−64) seed 1Δ(512−64) seed 2sign stable
Librispeech-PC6.25+1.59+4.55+6.18yes
Librispeech-PC2−2.15−1.89−1.68yes
E2 TTS6.25+1.96+4.38+6.03yes
E2 TTS2−1.68−1.71−1.60yes
SeedTTS6.25+4.78+7.73+8.83yes
SeedTTS2−2.95−2.88−2.64yes

Please report training-run variability

WER (%) on LibriSpeech-PC / E2 TTS suite / Seed-TTS EN. Three generators per cell were trained independently from scratch (seed 0 is the run reported in the main text; seeds 1–2 are retrained under an identical tokenizer, data and schedule), and each was rendered with three flow-decoder seeds. All seed entries are average over three render seeds. SD(train) is the standard deviation across the three training runs; SD(inference) is the within-run standard deviation across the three render seeds. Cells are named token rate × code height. These consistent results show that our numbers are not just a single lucky train run.

cellseed 0seed 1seed 2SD(train)SD(inference)
6.25 × 641.74 / 1.81 / 1.871.71 / 1.75 / 1.771.75 / 1.77 / 1.790.02 / 0.03 / 0.050.008 / 0.023 / 0.059
2 × 645.26 / 5.65 / 7.405.04 / 5.50 / 7.205.01 / 5.47 / 7.210.14 / 0.10 / 0.110.118 / 0.085 / 0.161
2 × 5123.11 / 3.97 / 4.453.15 / 3.79 / 4.333.32 / 3.87 / 4.580.11 / 0.09 / 0.130.050 / 0.063 / 0.160

and confidence intervals for the WER differences.

Confidence intervals on the WER difference between code heights. d = WER(h=512) − WER(h=64) in WER points, so d > 0 means the wide code is worse. Each d(seed k) is one independently trained generator, already averaged over its three flow-decoder seeds. Two intervals are reported because two sources of variance are at issue: the across-training-runs interval is a t-interval on the mean of the three d values (df = 2, t* = 4.303), and the across-utterances interval is a paired bootstrap over test utterances (4,000 resamples, pooled over all nine runs per cell). Positive and negative d indicate opposite height preferences, so the sign of d is the quantity our conclusion depends on. This sign is identical in all 18 cells.

Eval setRated(seed 0)d(seed 1)d(seed 2)mean d95% CI across training runs95% CI across utterances
LibriSpeech-PC6.25 tok/s+1.59+4.55+6.18+4.11
−1.68,+9.89−1.68, +9.89

|

+3.8,+4.4+3.8, +4.4

| | E2 TTS suite | 6.25 tok/s | +1.96 | +4.38 | +6.03 | +4.12 |

−0.96,+9.21−0.96, +9.21

|

+3.8,+4.5+3.8, +4.5

| | Seed-TTS EN | 6.25 tok/s | +4.78 | +7.73 | +8.83 | +7.11 |

+1.91,+12.32+1.91, +12.32

|

+6.6,+7.7+6.6, +7.7

| | LibriSpeech-PC | 2 tok/s | −2.15 | −1.89 | −1.68 | −1.91 | [−2.48, −1.33] | [−2.1, −1.7] | | E2 TTS suite | 2 tok/s | −1.68 | −1.71 | −1.60 | −1.66 | [−1.80, −1.52] | [−1.9, −1.5] | | Seed-TTS EN | 2 tok/s | −2.95 | −2.88 | −2.64 | −2.82 | [−3.23, −2.41] | [−3.2, −2.5] |

A-Q4. What measurable benefit comes specifically from joint configurability, beyond the benefits of low-rate representations and reusing a tokenizer checkpoint, given that downstream models are trained separately for different configurations?

The main measurable practical benefit is the efficiency of encoder training. With Laconic, one encoder is with 20.8 GPU-hours and serves all 36 cells of a grid (6 rates × 6 widths). In contrast, training all 36 versions separately would cost ≈639 GPU-hours, roughly 30 times more .

A secondary benefit is storage efficiency. To be able to consider embeddings with different widths, with Laconic, one only needs to store the maximal width (512), since shorter width embeddings are prefixes of it. In contrast, with a separate model the different width embeddings are independent and must be stored separately, nearly doubling the storage footprint.

Critical Weaknesses:

A-W1. The incremental contribution and competitiveness against relevant continuous representations are insufficiently established.
Several uncited works directly overlap with the paper’s methodological motivation: Semantic-VAE and HoliTok investigate semantic alignment and task-informed representation learning; WavCube and LoSATok study compact semantic–acoustic representations using feature reconstruction. These works provide relevant comparison points for reconstruction, understanding, and generative modeling. However, the paper neither clearly distinguishes its incremental contribution from these approaches nor directly compares representative LACONIC configurations with them under appropriately controlled conditions.

Our motivation differs in two main ways from this prior work (though we will cite and explain -- thanks for pointing this out!):

  1. Number of tokens per second:

All of the models you mention are in the high tokens/sec regime (Semantic-VAE is 40 tokens/sec, HoliTok is 25 tokens/sec, LoSATok is 25 tokens/sec, WavCube is 50 tokens/sec). We aim at far lower token rates (6.25 to 1 tokens/sec).

  1. Per-task choice of dimensionality:

All of the cited models use a single dimensionality. This may be suboptimal, as the best dimensionality for one task, say ASR, is not necessarily the same as for another task, say TTS. We train a single model from which we can extract representations with different dimensionalities depending on the task.

A-W2: The evidence does not sufficiently support the generality and reliability of the claimed trade-offs.

We believe we have addressed the details of this section in our responses to your questions.

I appreciate the authors' careful efforts to address most of my concerns. Although the writing still need to be polished (The authors could improve readability by focusing more on the key findings and overall trends rather than reiterating individual numerical results already presented in the tables), I have decided to raise my score from 2 to 3. Looking forward to the public release of the code.

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

The paper studies how the capacity budget of a continuous speech representation should be split between token rate and token dimensionality rather than described by token rate alone. It introduces LACONIC, a single encoder distilled from a frozen WavLM-Large and trained with a nested objective, so that token rate is selectable at encoding time and token dimensionality is selectable at reading time as a nested prefix without re-encoding. Across ASR probing, a speaker-identity probe, reconstruction, and zero-shot TTS, the authors report that rate and dimensionality partially substitute in a mid range, that a temporal floor appears at 1 token per second, and that content, speaker identity, and quality occupy different regions of the rate-dimension space.

Critical strengths
  1. The reframing of representation capacity as a two-dimensional rate-times-dimension allocation exposed by one encoder with inference-time control over both axes is a clean and useful conceptual contribution.
  2. The empirical characterization is thorough and honest, giving full rate-dimension surfaces for ASR, reconstruction, and TTS plus ablations that test whether the frontier is an artifact of training setup.
Critical weaknesses
  1. The whole study is trained and evaluated on read English audiobook speech, so the temporal floor and the rate-dimension exchange may be artifacts of clean low-density speech, with noisy and multilingual data left in appendices.
  2. The representation is a distillation of a frozen WavLM-Large and inherits its ceiling, so the contribution is a reordering of existing features rather than a representation learned from audio.
Questions for authors
  1. Does the temporal floor at 1 token per second persist for higher-density speech such as fast, spontaneous, or non-English speech?
  2. Since the utterance tokens carry almost all speaker identity yet are excluded from the budget, how do the trade-off conclusions change 3. if they are counted, and is five tokens tuned or arbitrary? How much of the reported geometry is specific to the WavLM teacher and the layer-6 target rather than a general property of speech?
1 comment

Questions For Authors:

A-Q1. Does the temporal floor at 1 token per second persist for higher-density speech such as fast, spontaneous, or non-English speech?

Yes. The table below shows word error rates (%) for reconstruction and ASR, showing that a similar pattern holds both for spontaneous speech from Emilia (which contains naturally fast spoken utterances) and for VoxPopuli-FR (French, i.e. non-English speech).

TaskOriginal audioWavLM at 50 tok/sLaconic at 6.25 tok/sLaconic at 5 tok/sLaconic at 4 tok/sLaconic at 3.13 tok/sLaconic at 2 tok/sLaconic at 1 tok/s
Emilia-EN (spontaneous)Reconstruction7.5—7.97.98.08.311.659.7
ASR probe—9.712.813.314.015.720.944.5
VoxPopuli-FR (non-English / French)Reconstruction8.5—9.89.910.010.815.891.5
ASR probe—20.923.824.925.928.537.060.3

A-Q2. Since the utterance tokens carry almost all speaker identity yet are excluded from the budget, how do the trade-off conclusions change if they are counted, and is five tokens tuned or arbitrary?

Since the utterance tokens carry almost all speaker identity yet are excluded from the budget, how do the trade-off conclusions change if they are counted

For any given speech utterance, the utterance-level token budget is a constant, and thus doesn’t change the trade-off conclusions.

and is five tokens tuned or arbitrary?

In preliminary experiments we tried 1, 5 and 10 utterance tokens. We found 5 to be better than 1 and similar to 10, and therefore use 5.

A-Q3. How much of the reported geometry is specific to the WavLM teacher and the layer-6 target rather than a general property of speech?

We are not quite sure what is meant by the geometry being a general property of speech. If you are asking whether it would hold for other learned representations besides WavLM: We do not claim that our findings hold beyond the WavLM teacher. We suspect that the findings would be similar for other commonly used encoders that are trained with similar objectives, but we have not verified this and do not claim it.

To clarify why we chose WavLM and layer 6:

Why the WavLM teacher

We chose to use WavLM and not Whisper or Mimi as teacher since in early experiments the latter produced worse results (see Table 4 in Appendix C.5 for Whisper, and Table 21 in Appendix H for Mimi)

Why the layer 6 target

As stated in L700-L704, layer 6 was chosen as it has been reported from previous studies that it is best for speaker-preserving synthesis and could also inherit a free off-the-shelf vocoder.

Critical Weaknesses:

A-W1. The whole study is trained and evaluated on read English audiobook speech, so the temporal floor and the rate-dimension exchange may be artifacts of clean low-density speech, with noisy and multilingual data left in appendices.

Addressed in the answer to Q1

A-W2. The representation is a distillation of a frozen WavLM-Large and inherits its ceiling, so the contribution is a reordering of existing features rather than a representation learned from audio.

We disagree that the representation inherits the frozen WavLM-Large ceiling. The way in which an encoder “organizes” information affects its performance, even though it does not add information. For example, we can see in Figure 4 (right panel) that Laconic features have substantially better speaker accuracy than WavLM (98.4% vs 88.4%).

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.