acceptodds
Under review as a conference paper at ICLR 2027

MyoCodec: A Streaming Neural Codec for Electromyography

Abstract

Neural codecs encode continuous signals into compact sequences of discrete tokens, providing an interface for efficient transmission, storage, and token-based sequence modeling. This paradigm has been widely adopted in modern speech and audio frameworks; however, the biosignal domain still lacks a neural codec designed specifically for low-bitrate streaming and generalization across diverse downstream tasks. We present MyoCodec, a streaming neural codec designed for electromyography (EMG). Inspired by recent neural audio codecs, MyoCodec combines causal Transformers with residual vector quantization to encode continuous EMG signals into different levels of EMG representations spanning from continuous latent features to discrete tokens operating at 50 Hz. Trained on twelve public EMG datasets, MyoCodec achieves favorable performance in both intrinsic codec quality and representative downstream tasks, including typing (emg2qwerty), hand-pose (emg2pose), speech decoding (emg2speech), and speech-to-EMG synthesis (speech2emg). Across these tasks, MyoCodec exhibits strong performance against prior models while providing a compact and causal EMG representation. During streaming inference, it requires compute time of only 0.482 ms for each 20 ms frame, enabling real-time streaming. Also, the discrete token representation provided by MyoCodec has the potential to support integration into language-model based approaches, creating a path toward LLM-based interactive systems, where tokenized EMG representations are directly processed into such language or speech models. Code and model weights will be released publicly upon publication.

Primary area:
applications to neuroscience & cognitive science
Keywords:
neural codec, electromyography, EMG, streaming

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_tARpOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

The paper presents MyoCodec, a causal neural codec for surface EMG. It uses an encoder–RVQ–decoder design adapted from a recent streaming speech codec (JHCodec) and the training losses of BioCodec. Each EMG channel is coded independently. The 2 kHz signal is cut into 40-sample patches (50 Hz frames), passed through an 8-layer causal Transformer with a 64-frame sliding window, quantized with a 6-stage RVQ (256 entries, 32-d), and decoded by a mirrored causal Transformer. The total bitrate is 2400 bit/s per channel and the model has 12.7 M parameters. Training has two phases: reconstruction plus VQ losses for 100k steps, then a linearly ramped MPD/MS-STFT adversarial and feature-matching phase for another 100k steps.

Critical strengths
  1. Timely and well-scoped problem. Neural-codec tokenization is established for speech and audio but not for EMG. A streaming, causal, compact EMG tokenizer is a reasonable and useful thing to build, especially for silent-speech and wristband HCI.
  2. Broad downstream coverage. Four tasks covering discrete recognition (typing), continuous regression (pose), cross-modal generation EMG to speech and speech to EMG is more diverse than most biosignal-tokenizer papers. On emg2pose and emg2speech, MyoCodec improves over all front ends in the authors' setup.
  3. Real streaming implementation. The paper ships a KV-cached, frame-wise implementation with CUDA Graphs and reports per-frame encoder and decoder compute on both GPU and CPU. Many "causal" codec papers never measure this.
Critical weaknesses
  1. The baseline comparisons are confounded, so the gains cannot be attributed to MyoCodec's design. The confound is that MyoCodec and the baselines are trained on different data with different preprocessing. MyoCodec uses 12 corpora covering about 24k channel-hours. The released BioCodec was trained on emg2qwerty only, with band-pass filtering. (For emg2speech (Table 5) the authors did retrain BioCodec, but the paper never says on what data. That comparison might be partly controlled, or might not be.)

  2. The "codec / discrete token" thesis is largely untested downstream.

    • Three of four tasks bypass quantization. emg2pose uses the pre-quantization encoder output for all front ends (L892–893). emg2speech uses frozen continuous features (L900). speech2emg predicts the continuous bottleneck projection (L913–914). Only emg2qwerty uses the quantized output, and even there it is the dequantized embedding sum fed to a continuous network (L877–878), not discrete tokens.
    • No token-level modelling experiment. There is no next-token prediction over MyoCodec tokens, no token-based classifier or LM, no token-based speech2emg, and no downstream-performance-vs-Nq curve. Yet the abstract , intro, Section 3.3 and the conclusion all lean heavily on LLM integration.
  3. Heavily missing ablations. DAC is a good example of what's missing. The paper cites SoundStream, EnCodec, DAC, Mimi and TS3-Codec only in the introduction and related work. None is trained on EMG or compared against. These aspects are also missing, e.g., Does the Transformer backbone help? Is the data or the model doing the work? Does adversarial training help? ...

Questions for authors
  1. Which version of BioCodec (released or retrained) is used in Tables? On what data, and with what preprocessing, was the retrained BioCodec of Table 5 trained?
  2. Why choose 50Hz frame rate? Have you tested others?
2 comments

We thank the reviewer for constructive review.

W1. (Also stated in our response to vmup) Comparison against BioCodec. We compare systems, each a model together with its pretraining data and preprocessing, and assembling a pretraining corpus from twelve public EMG datasets is part of MyoCodec's contribution. We do not claim that MyoCodec's architecture outperforms BioCodec's independently of the training data. Comparing codecs pretrained on different datasets is also common practice for neural audio codecs, as in JHCodec (Lee, 2026).

MyoCodec is, to our knowledge, the first streaming neural codec for EMG, so no prior codec can be compared with it under matched conditions. We chose BioCodec for comparison because it has shown promise on EMG despite being primarily developed for EEG. The comparison is one proxy for where MyoCodec stands, not an absolute ranking. A further contribution of the paper is the release of MyoCodec's model weights and source code, so that future work can build a controlled baseline.

W2. We ran additional experiments comparing token and continuous feature input for MyoCodec (refer to our general response). As stated above, controlled comparison of token and continuous-feature inputs is challenging to design. Existing downstream models were developed for continuous inputs, and discrete tokens may pay off only with substantial architectural changes, such as modeling the first codebook separately from the residual ones. This also happened in the speech domain where speech modeling paradigm shift happened after the rise of neural audio codecs (e.g., AudioLM [Borsos et al., 2023], VALL-E [Wang et al., 2023], MusicGen [Copet et al., 2023]). The token vs continuous input results should therefore be interpreted with caution. More broadly, we present continuous features and tokens as options to choose from by task and use case, not as competitors. The core contribution is the idea of a streaming neural codec for EMG, which is, to our knowledge, the first.

W3. Our core contribution is the first streaming neural codec for EMG. We agree that further ablations, such as other codec architectures trained on EMG and the Transformer backbone, would strengthen the paper, and we leave them for future work. On adversarial training, we ran the ablation the reviewer asks for; the results are below and in Table 8 in Appendix E. The results suggest the efficacy of adversarial training.

ModelSI-SDR (dB)SNR (dB)rwavr_\mathrm{wav}renvr_\mathrm{env}
End of Phase 17.297.590.8970.968
+ 100k steps without the discriminator7.007.500.8920.967
+ 100k steps with it (reported model)9.709.970.9320.967

Q1. (Also stated in our response to the other reviewers) Tables 1, 3, 4 and 6 use the official BioCodec's EMG model, pretrained on emg2qwerty. For Table 5 we retrained BioCodec on emg2qwerty with a wider band (20–850 Hz) and a higher frame rate of 50 Hz, as its original version did not converge during training, potentially due to the frequency range crucial for speech-related information in EMG signals. We also revised the table note in the updated pdf.

Q2. We chose 50 Hz mainly because it is a common frame rate for neural audio codecs and the operating frame rate of the backbone architecture we build on. Early in this work, we compared several frame rates and found that reconstruction quality increased with the frame rate. These results come from early experiments with an architecture that differs from the final one, so they should be interpreted with caution.

Frame rateSI-SDR (dB) ↑r_wav ↑R² ↑NRMSE ↓MAE ↓r_env ↑
12.5 Hz2.690.7970.6300.6080.3980.948
25 Hz5.290.8670.8040.4420.2380.957
50 Hz10.070.9340.9030.3120.1490.969

I thank the authors for the response and the revision. I will maintain my score because both issues remain, that baselines are all under different setups to establish fair, obvious comparisons, while simultaneously the ablations are still incomplete. I also encourage the authors to center the paper around an enhanced central claim over "the first streaming neural codec for EMG".

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

This paper presents MyoCodec, a streaming neural codec for electromyography. It combines a channel-independent causal Transformer encoder–decoder with residual vector quantization, producing representations at a frame rate of 50 Hz and a full bitrate of 2,400 bits/s per channel. The model is pretrained on twelve public EMG datasets using reconstruction and quantization objectives, followed by an additional adversarial training stage. Evaluation covers waveform reconstruction, typing, hand-pose estimation, speech decoding, and speech-to-EMG synthesis. The downstream experiments use quantized features for typing and continuous features or prediction targets for the other tasks. The paper also reports cached frame-wise inference on both GPU and CPU.

Critical strengths
  1. MyoCodec combines low-bitrate EMG compression with frame-by-frame processing. With two codebooks at 800 bits/s per channel, it outperforms BioCodec and TinyMyo on all four reconstruction metrics in Table 1. The reported encoding and decoding times total 0.482 ms on GPU and 6.18 ms on a single CPU thread per 20 ms frame, although the channel count for these measurements is not specified. The compression quality and streaming implementation make this a useful contribution for EMG interfaces.
  2. The experiments cover typing, hand-pose estimation, speech decoding, and EMG synthesis. Several comparisons use common downstream components to evaluate different EMG representations. MyoCodec achieves a speech-decoding WER of 21.12%, compared with 27.77% for TinyMyo and 28.33% for the retrained BioCodec. Its hand-pose joint-angle error is 14.24°, compared with 14.75° for BioCodec and 14.83° for TinyMyo. These results suggest that the pretrained model provides useful features for different EMG tasks, including when its continuous representations are used.
Critical weaknesses
  1. The comparison with BioCodec is difficult to assess because the pretraining conditions differ. MyoCodec is trained on twelve EMG datasets, while the BioCodec EMG model described in Appendix A was pretrained on emg2qwerty. BioCodec is retrained for speech decoding, but the retraining data and setup are not clearly described. The paper does not report a comparison with matched pretraining data and preprocessing, and the evaluation data and aggregation procedure for Table 1 are also unclear. Since the method builds on existing codec designs, I do not find the current comparisons sufficient to establish how much the proposed changes improve performance under comparable training conditions.
  2. The usefulness of the low-bitrate representation across tasks remains unclear. Appendix D states that emg2qwerty uses quantized RVQ outputs, while emg2pose and emg2speech use continuous features and speech2emg predicts continuous latents. The authors clearly disclose these choices. The typing results provide evidence that the quantized features are useful, but there is no comparison showing whether the gains on pose and speech tasks are retained after quantization. This leaves the downstream usefulness of the low-bitrate representation established mainly for typing, which limits the evidence for its broader use across EMG tasks.
Questions for authors
  1. Which held-out datasets and aggregation procedure were used for Table 1, and were all models evaluated against identical target waveforms after the same resampling and filtering? What data and preprocessing were used to retrain BioCodec for speech decoding?
  2. What channel count and batching configuration were used for the reported per-frame compute times of 0.482 ms on GPU and 6.18 ms on CPU? Are CPU measurements available for the full 16-channel emg2pose and 32-channel emg2qwerty configurations?
  3. How many random seeds were used for Tables 3–5? If multiple runs are available, please report their means and standard deviations, particularly for the small CER differences in Table 3.
  4. A comparison with BioCodec pretrained on the same data and preprocessing, with a comparable training budget, would help establish the benefit of the proposed codec design. Could you provide such a comparison or any existing results that control for these differences?
  5. Could you compare continuous and quantized RVQ features on emg2pose or emg2speech while keeping the downstream model and training setup fixed? Results on either task would help establish whether the reported gains are retained when using the low-bitrate representation.
2 comments

We thank the reviewer for the constructive review.

W1/Q4. Comparison against BioCodec. We compare systems, each a model together with its pretraining data and preprocessing, and assembling a pretraining corpus from twelve public EMG datasets is part of MyoCodec's contribution. We do not claim that MyoCodec's architecture outperforms BioCodec's independently of the training data. Comparing codecs pretrained on different datasets is also common practice for neural audio codecs, as in JHCodec (Lee, 2026).

MyoCodec is, to our knowledge, the first streaming neural codec for EMG, so no prior codec can be compared with it under matched conditions. We chose BioCodec for comparison because it has shown promise on EMG despite being primarily developed for EEG. The comparison is one proxy for where MyoCodec stands, not an absolute ranking. A further contribution of the paper is the release of MyoCodec's model weights and source code, so that future work can build on top of it.

W2/Q5. We ran additional experiments comparing token and continuous feature input for MyoCodec (refer to our general response). As stated above, controlled comparison of token and continuous-feature inputs is challenging to design. Existing downstream models were developed for continuous inputs, and discrete tokens may pay off only with substantial architectural changes, such as modeling the first codebook separately from the residual ones. This also happened in the speech domain where speech modeling paradigm shift happened after the rise of neural audio codecs (e.g., AudioLM [Borsos et al., 2023], VALL-E [Wang et al., 2023], MusicGen [Copet et al., 2023]). The token vs continuous input results should therefore be interpreted with caution. More broadly, we present continuous features and tokens as options to choose from by task and use case, not as competitors. The core contribution is the idea of a streaming neural codec for EMG, which is, to our knowledge, the first.

Q1. (Also in our reply to reviewer U8Nd) Table 1 scores single-channel 5 s windows from the validation set of the pretraining corpus. Scores are averaged over windows and the validation set of each corpus is listed in Table 11 (Appendix G). We released the evaluation script used for the table, and we revised the manuscript to reflect this information. We additionally report reconstruction quality against matched common reference signals: the 2 kHz signal, the 1 kHz signal and the 0–500 Hz band (below and Table 9 in Appendix F in the revised pdf). MyoCodec outperforms BioCodec in all settings, including at a lower bitrate.

SI-SDR / SNR (dB), rwavr_\mathrm{wav} / renvr_\mathrm{env}2 kHz reference1 kHz reference0–500 Hz band
BioCodec (1,333 b/s)−1.73 / 2.31, 0.633 / 0.888−1.51 / 2.35, 0.641 / 0.889−1.44 / 2.38, 0.644 / 0.890
MyoCodec, 2 codebooks (800 b/s)3.51 / 4.83, 0.830 / 0.9303.92 / 5.11, 0.842 / 0.9303.96 / 5.12, 0.843 / 0.931
MyoCodec, 6 codebooks (2,400 b/s)9.70 / 9.97, 0.932 / 0.96710.57 / 10.75, 0.942 / 0.96810.63 / 10.78, 0.942 / 0.968

Tables 1, 3, 4 and 6 use the official BioCodec's EMG model, pretrained on emg2qwerty. For Table 5 we retrained BioCodec on emg2qwerty with a wider band (20–850 Hz) and a higher frame rate of 50 Hz, as its original version did not converge during training, potentially due to the frequency range crucial for speech-related information in EMG signals. We also revised the table note in the updated pdf.

Q2. timing configuration. Both reported figures are one channel stream at batch size 1. Additional CPU timings at more channels, per 20 ms frame are below:

Threads1 channel8 channels16 channels (emg2pose)32 channels (emg2qwerty)
16.18 ms11.51 ms18.01 ms29.88 ms
244.24 ms6.35 ms7.55 ms9.46 ms

Q3. We first would like to respectfully note that our core contribution is presenting the first streaming neural codec for EMG, for which no completely controlled baseline exists, so these comparisons are reference points and not a ranking. We ran some additional experiments with different seeds and report below. We believe our core claims hold as long as MyoCodec performs on par with the baselines.

TableRow and metricSeedsMean ± std
Table 4 (emg2pose)MyoCodec tokens, joint-angle error, val / test313.19 ± 0.04 / 14.96 ± 0.08
Table 5 (emg2speech)MyoCodec tokens, test WER321.02 ± 0.98

Thank you for the clarifications and the extra experiments. The tokens work well for speech decoding, but the hand-pose test error goes up from 14.29° to 14.96° with quantization. That’s slightly higher than the errors for both BioCodec and TinyMyo. Given these mixed results, I’ll keep my original score.

Reviewer_U8NdOpenReview ↗
Rating
2
Confidence
1
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

MyoCodec adapts a causal Transformer codec with RVQ to surface EMG. It processes each electrode channel independently, turning a 2 kHz signal into a 50 Hz representation with a nominal bitrate of 2.4 kbps per channel. The model is pretrained on 12 public EMG datasets using reconstruction and adversarial objectives. The experiments cover reconstruction quality, a bitrate sweep, and four downstream tasks: typing, hand-pose estimation, EMG-to-speech, and speech-to-EMG. Depending on the task, the downstream model uses either quantized features or continuous features from the codec. The paper also measures codec inference speed on GPU and CPU and presents possible applications in wearable interfaces and token-based sequence modeling.

Critical strengths
  1. The paper addresses a useful problem and evaluates more than reconstruction alone. A compact EMG representation that can support typing, pose estimation, and speech-related tasks would be valuable. Using a shared per-channel codec also makes it straightforward to handle different channel counts. I like the breadth of the downstream evaluation, particularly the inclusion of both EMG-to-speech and speech-to-EMG.

  2. The reported quality and codec efficiency results are promising. The bitrate sweep suggests that the reconstruction gains are not simply due to using more bits: the two-codebook setting reportedly exceeds the baselines at 800 bps per channel. The pose and EMG-to-speech results also show useful gains, and the codec has an implemented frame-by-frame inference path with GPU and CPU timing results. Subject to the protocol concerns below, these would make a worthwhile empirical contribution.

Critical weaknesses
  1. I could not reconcile the data and reconstruction evaluation protocols with the released code. Section 2.3 says that codec pretraining uses only the official training partition of datasets used downstream. However, the released corpus-building script assigns emg2qwerty users to new splits by hashing their IDs. Applying that function to the eight held-out users in the official generic-model test configuration puts all eight in the training split. If the input directory contained the full dataset and there was no additional filtering, their recordings would therefore enter codec pretraining. There may be a separate filtering step or a different script behind the reported results, but the final training manifest is needed to resolve this. This matters especially for interpreting results on unseen subjects.

    The reconstruction comparison is also difficult to assess. The paper gives a corpus inventory and general split rules, but does not identify the dataset composition, sample counts, or aggregation weights behind Table 1. In tools/eval_recon.py, MyoCodec is evaluated against the 2 kHz shard signal, while BioCodec is evaluated against a version resampled to 1 kHz, standardized again, and clipped. These are different reference signals. Native-input evaluations can be useful, but they do not by themselves establish a common-reference rate–distortion comparison. I would need the actual evaluation commands, data lists, and reference-signal definitions to understand how much of the reported difference comes from the codecs.

  2. The evidence for real-time operation stops at the codec, and the end-to-end causal protocol is not yet clear. Table 7 reports codec compute time; the downstream tables report task quality and, in some cases, lookahead. I did not find downstream runtime, first-output latency, or a complete latency budget. For example, EMG-to-speech also needs the temporal predictor and speech decoder, while typing needs a policy for committing CTC/beam-search outputs. A bounded-lookahead network alone does not establish that these complete systems run in real time. The channel count and batch setting used for the timing measurements also need to be specified.

    There are concrete implementation details behind this concern. data/extract_shards.py normalizes each selected channel using statistics from the full five-second, multichannel window. How those statistics are obtained online is not explained. Table 5's uniform 100 ms lookahead description also conflicts with the offline frontend in App. D. Finally, the fast StreamingSession uses the reference path for its first 65 frames before switching to the CUDA Graph path, whereas tools/verify_streaming.py checks the ordinary streaming_step implementation. This does not imply a 1.3-second wait for the first output—the reference path returns frames immediately—but it does mean that startup timing and fast-path quality need their own checks. These gaps leave the connection between the reported quality, speed, and wearable-system claims incomplete.

Questions for authors
  1. What exactly is the reconstruction evaluation set in Table 1: which corpora, subjects, sessions, and how many windows from each? Are scores averaged over windows or first averaged within each dataset? Does the released MyoCodec/BioCodec evaluation path correspond to the table?

  2. What is the end-to-end future-context budget for each system in Tables 5 and 6, including preprocessing and the speech frontend/decoder where applicable? In particular, how should the bidirectional Gaddy and Klein frontend be read alongside Table 5's caption? How are normalization statistics obtained during online use, and, if reconstruction in the original signal units is intended, how does the receiver obtain the statistics needed to restore that scale?

  3. Can the reported task quality and speed be demonstrated together using the actual streaming implementation? For a representative downstream task, it would help to report the channel count, hardware, sustained real-time factor, and first-output latency, with startup and steady-state costs separated. The released training windows are five seconds long, and the verification script also uses short windows; longer continuous recordings would help check behavior once the stacked layers have accumulated more history.

  4. Which pose frontend produced the first row of Table 4, and what are the frontend parameter counts and pretraining conditions for the compared systems? The original public TDS configuration has about 2.63M frontend parameters, compared with about 6.3M for the MyoCodec encoder, but it is unclear whether that is the configuration used here. Keeping the pose decoder fixed is helpful; a matched-capacity or same-architecture initialization control would further clarify how much comes from codec pretraining rather than frontend size or architecture.

  5. In speech-to-EMG, is the predicted continuous bottleneck passed through RVQ before decoding, or fed directly to the decoder? Have you compared the same codec architecture with and without quantization, or continuous-feature regression with token prediction? These are not prerequisites for using a codec as a feature extractor, but such controls would help establish the contribution of the discrete representation beyond its familiar storage and transmission benefits.

  6. The phase term in losses.py applies L1 directly to torch.angle outputs. This treats phases near +π and −π as far apart even when their circular distance is small. Was this the loss used for the reported model? A comparison with a wrapped or cosine phase loss, and with no phase term, would help establish its effect. The current phase term is also unweighted by spectral magnitude, so low-energy bins may contribute unreliable phase targets.

1 comment

We thank the reviewer for the constructive review.

W1. We greatly appreciate the reviewer auditing the released code and finding the bug in our emg2qwerty preprocessing. As the reviewer pointed out, the original code split emg2qwerty randomly rather than by its official split. We fixed the bug and re-ran the experiments on the official split. The revised PDF reports the updated MyoCodec numbers, which show no substantial difference in overall trend, and Table 11 (Appendix G) shows the split of pretraining corpora. The updated code is available at the anonymized code link.

W2. Regarding real-time operation of the downstream tasks, we would like to respectfully note that the scope of this work is the codec itself. We adopted causal architectures for the downstream tasks, but they serve as examples that demonstrate MyoCodec's potential utility. We believe that proposing fully streaming downstream models for all four downstream tasks is beyond the scope of this work. We acknowledge this as a limitation in the revised manuscript. The timings in Table 7 are for one channel stream at batch size 1. For how the normalization statistics are obtained, see Q2.

Q1. Table 1 scores single-channel 5 s windows from the validation set of the pretraining corpus. Scores are averaged over windows and the validation set of each corpus is listed in Table 11 (Appendix G). We released the evaluation script used for the table, and we revised the manuscript to reflect this information. We additionally report reconstruction quality against matched common reference signals: the 2 kHz signal, the 1 kHz signal and the 0–500 Hz band (below and Table 9 in Appendix F in the revised pdf). MyoCodec outperforms BioCodec in all settings, including at a lower bitrate.

SI-SDR / SNR (dB), rwavr_\mathrm{wav} / renvr_\mathrm{env}2 kHz reference1 kHz reference0–500 Hz band
BioCodec (1,333 b/s)−1.73 / 2.31, 0.633 / 0.888−1.51 / 2.35, 0.641 / 0.889−1.44 / 2.38, 0.644 / 0.890
MyoCodec, 2 codebooks (800 b/s)3.51 / 4.83, 0.830 / 0.9303.92 / 5.11, 0.842 / 0.9303.96 / 5.12, 0.843 / 0.931
MyoCodec, 6 codebooks (2,400 b/s)9.70 / 9.97, 0.932 / 0.96710.57 / 10.75, 0.942 / 0.96810.63 / 10.78, 0.942 / 0.968

Q2. In Table 5, only the downstream model after the EMG features is causal. The Gaddy and TinyMyo rows extract their EMG features non-causally and then feed them to the causal downstream model, whereas BioCodec and MyoCodec extract their features causally. Likewise, only the MyoCodec row use causal running normalization; the other rows use 5 s block statistics. We revised the manuscript to make this clear.

Q3. As noted above, the scope of this work is the codec, and the downstream tasks serve to show where the proposed codec stands. We respectfully argue that developing fully streaming downstream models for each task are beyond this scope and are left for future work.

Q4. Table 4's first row is emg2pose's TDS network, the configuration the reviewer cites (2.63 M parameters), trained from scratch together with the pose decoder. The other three front ends are pretrained and frozen. Each of those systems trains about 4.4 M parameters, against 5.98 M for the TDS row, so trainable capacity does not explain MyoCodec's lead. We revised the manuscript for clarity.

As suggested, we ran the same-architecture control the reviewer suggests. With the MyoCodec encoder re-initialized at random and kept frozen, emg2pose error rises from 12.77° / 14.29° to 16.81° / 18.35° (val / test).

Q5. In the previously submitted version, the predicted continuous bottleneck latent is fed directly to the MyoCodec decoder, without passing through the RVQ (the continuous rows of Table 6). As noted in the general response, we added results with discrete tokens as targets (below), and updated the PDF.

speech2emg (Table 6), Env. CC at 120 ms50 ms window20 ms window
MyoCodec, continuous latent0.6500.576
MyoCodec, discrete tokens0.6040.524

Q6. We intended to follow BioCodec's training objective, which reports that this term keeps the model from drifting toward inverted waveforms. We agree a wrapped variant could be another option but we would leave it for future work.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.