- 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
- 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.
- 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.
- 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
-
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.)
-
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.
-
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
- 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?
- Why choose 50Hz frame rate? Have you tested others?
2 comments
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".
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.
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.