Pith. sign in

REVIEW 4 major objections 6 minor 25 references

HH-Codec: High Compression High-fidelity Discrete Neural Codec for Spoken Language Modeling

T0 review · 4 major / 6 minor · reviewed 2026-08-15 · deepseek-v4-flash

Pith's one-line read HH-Codec compresses speech to 24 tokens per second at 0.3 kbps while keeping reconstruction quality close to codecs using 3–9 kbps.

desk verdict The 24-token/s operating point is real and worth attention, but the paper's own table contradicts its headline claim and the distillation mechanism that supposedly makes it work is underspecified. read the letter →

arxiv 2507.18897 v1 pith:FAK2WEZC submitted 2025-07-25 cs.SD cs.AIeess.AS

classification cs.SDcs.AIeess.AS
keywords neuralaudiocodecvectorquantizationsingle-quantizerinferencesemanticdistillationspokenlanguagemodelingspeechcompressioncodebookutilizationreconstruction
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

HH-Codec is a neural speech codec that claims to turn 24 kHz audio into 24 discrete tokens per second, a bandwidth of 0.3 kbps, using only one codebook at inference. That is roughly ten to thirty times fewer tokens per second than common neural codecs, which spread information across several parallel quantizer streams. The paper argues the compression works because its SLM-VQ quantization space is shaped for spoken language: the first codebook is trained to align with HuBERT semantic features while a second residual quantizer, used only during training, acts as a regularizer. An asymmetric architecture reconstructs a mel-spectrogram before the waveform, with dual supervision and progressive fine-tuning of a pretrained BigVGAN decoder. Across noisy, clean, and out-of-domain speech, the codec reports reconstruction scores on par with systems using 3–9 kbps, and its tokens lower the loss of a downstream audio language model faster than two prior tokenizers under identical settings.

What carries the argument

The central machinery is SLM-VQ, a vector-quantization space designed for spoken language modeling: it builds on the SimVQ idea of a frozen codebook with a learnable MLP, adds a second residual VQ layer used only during training as a regularizer, and replaces the straight-through estimator with the rotation trick to improve gradient flow and codebook utilization. The distillation term (Eq. 2) is a per-dimension log-sigmoid cosine loss between a linear projection of the first quantizer's output and HuBERT features; it carries semantic content into a token rate of roughly one to two tokens per phoneme. On the decoder side, the Audio-VQ-Mel-Audio path—a mel-spectrogram reconstruction head followed by a BigVGAN waveform generator—is what turns sparse semantic-ish tokens back into high-fidelity audio; it is initialized from a pretrained BigVGAN and fine-tuned only after the encoder and quantizer stabilize, with the two-stage objective in Eq. 7.

What would settle it

Train the full HH-Codec recipe with the distillation loss in Eq. 2 removed, leaving the two-layer SLM-VQ and all reconstruction losses intact. If the drop in UTMOS, STOI, and speaker similarity is within the run-to-run noise of the reported numbers, then semantic distillation is not the mechanism that makes 24 tokens per second viable; if the codec collapses or the downstream audio-LLM loss no longer decreases faster than WavTokenizer's, the claim is supported.

Watch

Extended reading notes

Core claim

HH-Codec's central claim is that a single-quantizer codec running at 24 tokens/s (0.3 kbps) can reconstruct speech with the fidelity of multi-quantizer codecs that use 1.1–9 kbps, and can do so while remaining more efficient for spoken-language-model training. The mechanism offered for this is semantic distillation into a compact codebook: a cosine loss (Eq. 2) pulls a linear projection of the first quantizer's output toward HuBERT features, so that semantic content survives at roughly one to two tokens per phoneme, while a training-only second quantizer and a strong decoder preserve the acoustic detail needed for reconstruction. On LibriTTS test-other, test-clean, and Seed-TTS-eval, HH-Codec reports UTMOS of 3.21, 3.61, and 3.33 respectively at 0.3 kbps, with STOI, V/UV F1, and speaker-similarity scores comparable to or above several baselines. In a downstream comparison, its tokens produce faster audio-LLM pretraining loss reduction than WavTokenizer and CosyVoice tokens under identical hyperparameters.

Load-bearing premise

The load-bearing premise is that pushing the first quantizer's outputs toward HuBERT speech-meaning features at only 24 tokens per second teaches the discrete token lookup table enough about language without erasing the speaker's voice and the acoustic fine detail the decoder must rebuild.

Editorial extensions

If this is right

  • Speech-language models could consume a single token stream at 24 tokens/s, removing the complexity of parallel quantizer streams and cutting sequence length by an order of magnitude relative to 300–900 tokens/s codecs.
  • Transmission and storage of speech would drop to 0.3 kbps for a 24 kHz signal, about ten to thirty times below current neural codec bitrates, if the reported scores reproduce on other data.
  • The audio-LLM training loss curve implies that downstream speech models can reach a given loss with fewer steps or fewer tokens, which, if stable across scales, lowers the compute cost of pretraining speech LLMs.
  • The codebook utilization results (94% at 8192 entries) suggest the single codebook is not collapsing, so larger codebooks or longer training could push quality higher without adding quantizer streams.
  • The semantic and acoustic content shares one stream at text-like granularity, creating a path toward unified text–speech token spaces for joint models.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Editorial inference: the paper never ablates the HuBERT distillation term by itself; the 'w/ Single SLM-VQ' ablation removes the second quantizer, not the distillation. A reader should not conclude from the ablations alone that semantic distillation is the key ingredient—the two-layer regularizer may carry much of the benefit.
  • Editorial inference: if the distillation is load-bearing, the same recipe should transfer to other self-supervised speech teachers and to non-speech audio such as music or ambient sound; testing those transfers would reveal whether the 24-token/s result is specific to HuBERT-aligned speech or a general property of the architecture.
  • Editorial inference: the reported metrics are reconstruction-oriented (UTMOS, STOI, speaker similarity) and do not directly measure whether downstream tasks such as speech translation, emotion conversion, or prosody preservation inherit the semantic alignment; task-level evaluation would be the sharper test of the spoken-language-modeling claim.
  • Editorial inference: because the decoder is initialized from a pretrained BigVGAN and then fine-tuned, part of the fidelity gain may come from the vocoder's prior knowledge rather than from the codec's quantization space; comparing against a from-scratch BigVGAN decoder would separate those contributions.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 6 minor

Summary. HH-Codec proposes an ultra-low-bitrate neural speech codec that operates at 24 tokens per second (0.3 kbps) for 24 kHz audio while using a single quantizer at inference. The method combines an asymmetric encoder-decoder architecture, a two-layer SLM-VQ with HuBERT-based semantic distillation, and a progressive dual-supervision training strategy. The paper reports reconstruction metrics on three datasets, ablations, codebook utilization, and a downstream audio-LLM training-loss comparison, and it makes the code publicly available.

Significance. If the reported results are reproducible, a single-quantizer codec at 24 tokens/s with reconstruction quality close to systems using 3-9 kbps would be a substantive contribution, bringing audio token rates closer to text token rates. The paper ships code, evaluates with external metrics (UTMOS, STOI, WavLM-based SIM) against official open baselines, and the ablations show large effects for dual supervision and progressive training, which supports the plausibility of the architecture. However, the semantic distillation mechanism that is supposed to make 24 Hz viable is not fully specified and is never isolated in an ablation, and the stated performance claims are stronger than Table 1 supports. The downstream spoken-language-modeling evidence is also too thin. With clarifications and additional experiments, the central idea could be publishable, but the current manuscript needs substantial revision.

major comments (4)
  1. [Section 4.3, Table 1] The statement that HH-Codec 'outperforms a model using ten times the bandwidth in UTMOS but also matches its STOI, V/UV F1, and SIM scores' is not supported by Table 1. For the most natural 3 kbps comparison target, Vocos, on LibriTTS test-clean HH-Codec has STOI 0.89 vs 0.93, V/UV F1 0.90 vs 0.94, and SIM 0.73 vs 0.81; on test-other the gaps are similar. The comparison target should be named explicitly, and the claim should be restricted to the metrics and datasets for which the comparison actually holds.
  2. [Section 3.3, Eq. (2)] The distillation loss as written is dimensionally ill-posed. VQ1 operates at approximately 24 Hz while HuBERT features are typically extracted at 50 Hz, so the vectors VQ1(:,d) and H(:,d) have different lengths and the cosine similarity in Eq. (2) cannot be computed without an unstated interpolation, truncation, or alignment step. The authors must specify the exact tensor shapes and the resampling/alignment procedure, otherwise the semantic-transfer mechanism is unverifiable.
  3. [Sections 3.3 and 4.4, Table 2] The paper never isolates the HuBERT distillation loss Ldistill in an ablation. The 'w/ Single SLM-VQ' variant in Table 2 removes the second VQ layer but retains distillation, and no row in Table 2 trains without Eq. (2). Since the paper attributes the feasibility of 24 tokens/s to semantic distillation, an ablation that omits Ldistill is essential to support the central mechanism claim.
  4. [Section 4.5, Figure 2] The downstream audio-LLM experiment is not described in sufficient detail to support the claim that HH-Codec is effective for spoken language modeling. The figure shows only training loss curves with no specification of the LLM architecture, tokenization details, training hyperparameters, or number of runs, and no downstream task metrics such as ASR accuracy or speaker similarity are reported. This evidence is too weak to substantiate the downstream effectiveness claim.
minor comments (6)
  1. [Introduction, Section 1] Observation ④ states that UTMOS decreases by 63% when the token rate drops below 30 per second, yet the paper claims high fidelity at 24 tokens per second; please clarify how HH-Codec overcomes this degradation beyond the qualitative statements in Section 1.
  2. [Table 1] The table values are formatted with extra spaces (e.g., '3 .36'); please provide a clean machine-readable rendering with consistent decimal formatting.
  3. [Eq. (2)] Please clarify whether the cosine similarity is computed over the time dimension or the feature dimension and define the dimensions of VQ1 and H explicitly.
  4. [Table 3] Table 3 reports codebook utilization only; it would be informative to report reconstruction quality metrics for different codebook sizes so the reader can judge whether the utilization gains translate into fidelity.
  5. [Section 4] The paper states that LibriSpeech train-clean 100/360, VCTK, and an Emilia subset are used for training, but does not specify the mixing proportions or the exact composition of the final training set; this is relevant for interpreting the out-of-domain Seed-TTS-eval results.
  6. [Figure 2] Single loss curves without error bars or multiple seeds are difficult to interpret; please provide variance information or additional training runs.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: HH-Codec's headline reconstruction scores are evaluated with external metrics against open-weight baselines, and no central claim reduces by construction to a fitted input or a self-citation.

full rationale

Walking the derivation chain, the central claim is that a single-quantizer codec at 24 tokens/s reconstructs speech with objective quality comparable to much higher-rate codecs. The evaluation uses UTMOS, STOI, V/UV F1, and WavLM-based SIM, none of which appears in any training loss, and the baselines are scored from their official open-source weights, so the headline comparison is not fitted to the evaluator. The SLM-VQ construction (Eq. 1 VQ loss, Eq. 2 HuBERT cosine distillation, SimVQ prior, and the rotational trick) is a set of training supervisions; Eq. 2 uses an external HuBERT teacher that is not the evaluation metric, so the reported reconstruction scores do not reduce to the distillation objective by construction. The asymmetric Audio-VQ-Mel-Audio decoder and progressive training are architecture choices whose effects are measured by held-out reconstruction metrics, not by the training losses themselves. The downstream audio-LLM experiment in Fig. 2 is an empirical loss comparison under fixed model architecture and hyperparameters; even if comparing tokenizers at different token rates introduces a confound, that is an experimental validity concern, not circularity. There is no load-bearing self-citation chain: the cited priors (SimVQ, rotation trick, BigVGAN, WavTokenizer, HuBERT) are external works, and no 'uniqueness theorem' or prior-author result is invoked to force the design. The notable weaknesses are that the HuBERT distillation is not isolated in an ablation and Eq. 2's frame-rate alignment between 24 Hz VQ1 features and 50 Hz HuBERT features is unspecified; these are evidence and correctness gaps, not instances of a result being equivalent to its inputs by definition. Therefore the paper is not circular, and any concerns belong under correctness risk rather than circularity.

Assumptions & free parameters 6 free parameters · 4 assumptions · 0 invented entities

The system is entirely empirical; the reader pays for pretrained teachers (HuBERT, BigVGAN), prior VQ formulations (SimVQ, rotation trick), and a stack of hand-chosen hyperparameters (strides, codebook size, loss weights, phase threshold). There are no invented entities with falsifiable handles. The listed free parameters directly set the 0.3 kbps operating point and the training dynamics the central claims rest on.

free parameters (6)
  • Encoder stride configuration (8, 8, 4, 4) = 24 tokens/s
    The stride configuration in Section 3.2 is the single design choice that fixes the 0.3 kbps operating point; chosen by hand, not derived.
  • Codebook size = 8192
    Chosen for 0.3 kbps (12 bits per token). Utilization and downstream loss depend on it; the ablation shows sizes 1024-16384.
  • VQ loss commitment beta = 1.0
    Eq. 1, Section 3.3: set to 1 without sensitivity analysis.
  • Loss weights lambda_rec, lambda_D, lambda_distill, lambda_vq = not reported
    Eq. 7 defines the weighted total loss but the values are never given; central to the progressive training claims.
  • Phase-1 mel-loss threshold = 1.0
    Section 3.5: transition threshold for unfreezing BigVGAN; chosen by observation of training loss.
  • Training window length = longer than 1 s, exact value not specified
    Section 3.2 and the 'w/o Long windows' ablation show it matters (0.13 UTMOS drop), but the exact window is unstated.
assumptions (4)
  • domain assumption HuBERT representations are a valid semantic teacher for the single VQ layer
    Eq. 2 distills VQ1 output toward HuBERT features via cosine loss; if HuBERT is wrong-grained or misaligned, semantic retention at 24 tps is unsupported.
  • domain assumption Pretrained BigVGAN v2 (24 kHz, 100-band, 256x) can upsample the decoder's mel-spectrogram to high-fidelity audio without amplifying VQ errors
    Sections 3.4 and 3.5: final audio comes from frozen-then-fine-tuned BigVGAN; the whole fidelity story depends on this transfer.
  • domain assumption UTMOS predicted MOS and WavLM-based SIM correlate with human perception
    Section 4.2: all headline quality claims rest on these proxies; no human listening test.
  • ad hoc to paper The second RVQ layer trains as a 'Virtual Class' regularizer without changing inference
    Section 3.3: this is the paper's own mechanism to justify throwing away the second quantizer at inference; only indirectly supported by the single-SLM-VQ ablation.

how reviews work

0 comments
Cite this review

Pith. "Pith review of HH-Codec: High Compression High-fidelity Discrete Neural Codec for Spoken Language Modeling." pith.science (2026). https://pith.science/paper/FAK2WEZC

@misc{pith2026250718897,
  author       = {Pith},
  title        = {Pith review of: HH-Codec: High Compression High-fidelity Discrete Neural Codec for Spoken Language Modeling},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/FAK2WEZC}},
  note         = {Machine review of arXiv:2507.18897}
}
read the original abstract

Discrete speech tokenization is a fundamental component in speech codecs. However, in large-scale speech-to-speech systems, the complexity of parallel streams from multiple quantizers and the computational cost of high-time-dimensional codecs pose significant challenges. In this paper, we introduce HH-Codec, a neural codec that achieves extreme compression at 24 tokens per second for 24 kHz audio while relying on single-quantizer inference. Our approach involves a carefully designed Vector Quantization space for Spoken Language Modeling, optimizing compression efficiency while minimizing information loss. Building on this, we propose an asymmetric encoder-decoder architecture (Audio-VQ-Mel-Audio) that leverages dual supervision and progressive training to enhance reconstruction stability and fidelity. HH-Codec achieves state-of-the-art performance in speech reconstruction with an ultra-low bandwidth of 0.3 kbps. We further evaluate its effectiveness in codebook utilization and generative model adaptation, with extensive ablations validating the necessity of each module. HH-Codec is available at https://github.com/opendilab/HH-Codec.

Figures

Figures reproduced from arXiv: 2507.18897 by the authors.

Figure 1
Figure 1. The architecture overview of HH-Codec. Different color lines indicate the data flow used in inference and only for training. During inference, the audio is processed through the encoder and VQ1 to generate discrete quantization, which is then refined by the MLP. The decoder and fine-tuned BigVGAN subsequently reconstruct the Mel-spectrogram and audio. at inference. During training, the encoded output e of a 24 kHz i… view at source ↗
Figure 2
Figure 2. Efficient downstream audio-LLM training loss. 4.5. Downstream Task: Spoken Language Modeling To validate the effectiveness of our HH-Codec for spoken language modeling, we conduct ablation studies by inte￾grating this discrete codec with downstream large audio￾LLM training. Here we combine HH-Codec with Qwen2.5- 7B (Team, 2024). Our experiments systematically compare three state-of-the-art audio tokenization approac… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

25 extracted references · 4 canonical work pages

  1. [1]

    L., Almeida, D., Altenschmidt, J., Altman, S., Anadkat, S., et al

    Achiam, J., Adler, S., Agarwal, S., Ahmad, L., Akkaya, I., Aleman, F. L., Almeida, D., Altenschmidt, J., Altman, S., Anadkat, S., et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774,

  2. [8]

    Overview of the evs codec architecture

    Dietz, M., Multrus, M., Eksler, V ., Malenovsky, V ., Norvell, E., Pobloth, H., Miao, L., Wang, Z., Laaksonen, L., Vasi- lache, A., et al. Overview of the evs codec architecture. In 2015 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) , pp. 5698–5702. IEEE,

  3. [10]

    Textless speech emotion conversion using decom- posed and discrete representations

    Felix, K., Adam, P., Jade, C., Eugene, K., Tu-Anh, N., Wei- Ning, H., Abdelrahman, M., Emmanuel, D., and Yossi, A. Textless speech emotion conversion using decom- posed and discrete representations. arXiv preprint arXiv: 2111.07402,

  4. [12]

    Multi-scale sub- band constant-q transform discriminator for high-fidelity vocoder

    Gu, Y ., Zhang, X., Xue, L., and Wu, Z. Multi-scale sub- band constant-q transform discriminator for high-fidelity vocoder. In ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pp. 10616–10620. IEEE,

  5. [13]

    Emilia: An extensive, multi- lingual, and diverse speech dataset for large-scale speech generation

    He, H., Shang, Z., Wang, C., Li, X., Gu, Y ., Hua, H., Liu, L., Yang, C., Li, J., Shi, P., et al. Emilia: An extensive, multi- lingual, and diverse speech dataset for large-scale speech generation. In 2024 IEEE Spoken Language Technology Workshop (SLT), pp. 885–890. IEEE,

  6. [14]

    Wavtokenizer: an efficient acoustic discrete codec tokenizer for audio language modeling

    Ji, S., Jiang, Z., Wang, W., Chen, Y ., Fang, M., Zuo, J., Yang, Q., Cheng, X., Wang, Z., Li, R., et al. Wavtokenizer: an efficient acoustic discrete codec tokenizer for audio language modeling. arXiv preprint arXiv:2408.16532 ,

  7. [15]

    Bigvgan: A universal neural vocoder with large-scale training

    Lee, S.-g., Ping, W., Ginsburg, B., Catanzaro, B., and Yoon, S. Bigvgan: A universal neural vocoder with large-scale training. arXiv preprint arXiv:2206.04658,

  8. [16]

    Single-codec: Single-codebook speech codec towards high-performance speech genera- tion

    Li, H., Xue, L., Guo, H., Zhu, X., Lv, Y ., Xie, L., Chen, Y ., Yin, H., and Li, Z. Single-codec: Single-codebook speech codec towards high-performance speech genera- tion. arXiv preprint arXiv:2406.07422,

Show all 25 references
  1. [17]

    Finite scalar quantization: Vq-vae made simple

    Mentzer, F., Minnen, D., Agustsson, E., and Tschannen, M. Finite scalar quantization: Vq-vae made simple. arXiv preprint arXiv:2309.15505,

  2. [18]

    Enhanced direct speech-to- speech translation using self-supervised pre-training and data augmentation

    Popuri, S., Chen, P.-J., Wang, C., Pino, J., Adi, Y ., Gu, J., Hsu, W.-N., and Lee, A. Enhanced direct speech-to- speech translation using self-supervised pre-training and data augmentation. arXiv preprint arXiv:2204.02967 ,

  3. [19]

    arXiv preprint arXiv:2204.02152,

  4. [20]

    Neural machine translation of rare words with subword units

    Sennrich, R. Neural machine translation of rare words with subword units. arXiv preprint arXiv:1508.07909,

  5. [21]

    V ocos: Closing the gap between time-domain and fourier-based neural vocoders for high-quality audio synthesis

    Siuzdak, H. V ocos: Closing the gap between time-domain and fourier-based neural vocoders for high-quality audio synthesis. arXiv preprint arXiv:2306.00814,

  6. [22]

    H., Hendriks, R

    Taal, C. H., Hendriks, R. C., Heusdens, R., and Jensen, J. A short-time objective intelligibility measure for time- frequency weighted noisy speech. In 2010 IEEE inter- national conference on acoustics, speech and signal pro- cessing, pp. 4214–4217. IEEE,

  7. [25]

    J., Jia, Y ., Chen, Z., and Wu, Y

    Zen, H., Dang, V ., Clark, R., Zhang, Y ., Weiss, R. J., Jia, Y ., Chen, Z., and Wu, Y . Libritts: A corpus de- rived from librispeech for text-to-speech. arXiv preprint arXiv:1904.02882,

  8. [26]

    Addressing representa- tion collapse in vector quantized models with one linear layer

    Zhu, Y ., Li, B., Xin, Y ., and Xu, L. Addressing representa- tion collapse in vector quantized models with one linear layer. arXiv preprint arXiv:2411.02038,

  9. [2010]

    Seanet: A multi-modal speech enhancement network

    Tagliasacchi, M., Li, Y ., Misiunas, K., and Roblek, D. Seanet: A multi-modal speech enhancement network. arXiv preprint arXiv:2009.02095,

  10. [2015]

    Cosyvoice 2: Scalable streaming speech synthesis with large language models

    Du, Z., Wang, Y ., Chen, Q., Shi, X., Lv, X., Zhao, T., Gao, Z., Yang, Y ., Gao, C., Wang, H., et al. Cosyvoice 2: Scalable streaming speech synthesis with large language models. arXiv preprint arXiv:2412.10117,

  11. [2016]

    Neural codec language models are zero-shot text to speech synthesizers

    Wang, C., Chen, S., Wu, Y ., Zhang, Z., Zhou, L., Liu, S., Chen, Z., Liu, Y ., Wang, H., Li, J., et al. Neural codec language models are zero-shot text to speech synthesizers. arXiv preprint arXiv:2301.02111,

  12. [2018]

    Copet, J., Kreuk, F., Gat, I., Remez, T., Kant, D., Synnaeve, G., Adi, Y ., and D´efossez, A

    URL http://arxiv.org/ abs/1811.12611. Copet, J., Kreuk, F., Gat, I., Remez, T., Kant, D., Synnaeve, G., Adi, Y ., and D´efossez, A. Simple and controllable music generation. Advances in Neural Information Pro- cessing Systems, 36,

  13. [2020]

    Estimating or propagating gradients through stochastic neurons for con- ditional computation

    Bengio, Y ., L´eonard, N., and Courville, A. Estimating or propagating gradients through stochastic neurons for con- ditional computation. arXiv preprint arXiv:1308.3432 ,

  14. [2021]

    G., Duan, D., Iger, A., Liu, J

    Fifty, C., Junkins, R. G., Duan, D., Iger, A., Liu, J. W., Amid, E., Thrun, S., and R ´e, C. Restructuring vec- tor quantization with the rotation trick. arXiv preprint arXiv:2410.06424,

  15. [2022]

    Moshi: a speech-text foundation model for real-time dialogue

    D´efossez, A., Mazar ´e, L., Orsini, M., Royer, A., P ´erez, P., J ´egou, H., Grave, E., and Zeghidour, N. Moshi: a speech-text foundation model for real-time dialogue. arXiv preprint arXiv:2410.00037,

  16. [2023]

    Seed- tts: A family of high-quality versatile speech generation models

    Anastassiou, P., Chen, J., Chen, J., Chen, Y ., Chen, Z., Chen, Z., Cong, J., Deng, L., Ding, C., Gao, L., et al. Seed- tts: A family of high-quality versatile speech generation models. arXiv preprint arXiv:2406.02430,

  17. [2024]

    D´efossez, A., Copet, J., Synnaeve, G., and Adi, Y

    URL https: //arxiv.org/abs/2309.16588. D´efossez, A., Copet, J., Synnaeve, G., and Adi, Y . High fidelity neural audio compression. Transactions on Ma- chine Learning Research. D´efossez, A., Copet, J., Synnaeve, G., and Adi, Y . High fidelity neural audio compression. arXiv p...

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.