Pith. sign in

REVIEW 4 major objections 6 minor 20 references

A three-part detokenizer turns semantic audio tokens into high-fidelity speech at constant ~21 MB memory and ~16× real time on device.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · grok-4.5

2026-07-30 11:45 UTC pith:T7WQMN44

load-bearing objection Clean on-device systems advance: fully shared DiT-conditioned depth decoder plus fixed-window KV gives real constant-memory detokenization; the quality–memory conjunction is split across two configs. the 4 major comments →

arxiv 2607.23811 v1 pith:T7WQMN44 submitted 2026-07-26 cs.SD cs.CL

Memory Efficient Audio Synthesis with Decoupled Temporal Depth Diffusion Transformers

Dongseong Hwang , Prasanth Yadla , Kaan Elgin , Shifas Padinjaru Veettil , Sivanand Achanta , Dipjyoti Paul , Ramya Rasipuram , Tyler Johnson
show 3 more authors
Emad Soroush Chung-Cheng Chiu Zhifeng Chen
This is my paper
classification cs.SD cs.CL
keywords Speech SynthesisResidual Vector QuantizationDiffusion TransformersHigh-fidelity SynthesisOn-device detokenizerSliding-window attentionConstant memory
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

This paper presents the on-device audio detokenizer behind Siri Expressive Voices: a system that converts semantic tokens from a foundation model into residual vector quantization (RVQ) codes, then into speech, inside a tight shared memory budget. It splits the job into a streaming encoder, a temporal decoder, and one reusable depth decoder. The depth decoder generates every RVQ level with fully shared weights, guided only by Diffusion Transformer–style stage conditioning, instead of separate per-level decoders or projections. Causal sliding-window attention with a fixed-window key-value cache keeps runtime memory flat no matter how long the utterance is. On the Apple Matrix Coprocessor the design runs about 10 ms per step (~16× real time) at ~21 MB peak runtime memory, supports continuous 20–320 second streams, and, inside the full on-device stack, lifts Mean Opinion Score by +0.28 overall and +0.42 on conversational speech versus the prior production TTS.

Core claim

Decoupling temporal processing from depth (RVQ-level) processing, and identifying each RVQ level only through DiT-style stage conditioning on a single fully shared depth decoder, yields constant-memory, faster-than-real-time semantic-to-RVQ conversion that preserves synthesis quality under on-device constraints.

What carries the argument

The three-component detokenizer—streaming encoder, temporal decoder, and one reusable depth decoder with DiT-style stage conditioning (rotary encoding of the codebook index) plus fixed-window KV caching—carries the argument by making depth generation parameter-shared and memory independent of sequence length.

Load-bearing premise

That quality and speed measured on the research model still hold for the separately compressed production detokenizer, and that fixed-window attention plus mean RVQ feedback keep enough long-range context for high-fidelity streaming.

What would settle it

Run the shipped production detokenizer alone (same bitrate/frame rate as research) on held-out long utterances and check whether UTMOS/SI-SNR and ABX stay at or above the reported research numbers while peak runtime memory remains near 21 MB and independent of length out to several minutes.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Streaming expressive speech can run continuously on device beside a sparsely activated foundation model without memory growing with utterance length.
  • Per-level decoder stacks and per-codebook input/output projections become unnecessary once stage conditioning fully identifies RVQ depth.
  • Fixed-window KV caching makes 20–320 second synthesis practical under a shared DRAM budget of tens of megabytes.
  • Full-system MOS gains of about +0.3 overall and +0.4 on conversational speech become achievable at a 1B-parameter activation size on device.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same temporal–depth split and stage-conditioned shared decoder could transfer to other multi-codebook codecs beyond speech, wherever memory must stay flat under streaming.
  • If mean RVQ embedding feedback is the main temporal glue, replacing it with a learned temporal summary might further cut depth iterations without raising memory.
  • Constant-memory detokenizers lower the barrier to always-on multimodal assistants that keep audio generation resident while routing other experts in and out of DRAM.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 6 minor

Summary. The paper presents the detokenizer architecture behind Siri Expressive Voices: a three-module design (streaming encoder, temporal decoder, depth decoder) that converts semantic audio tokens from the AFM 3 Core Advanced foundation model into RVQ tokens. Its claimed contributions are (i) a single fully shared depth decoder that generates all RVQ levels autoregressively, identifying each level only via DiT-style stage conditioning (RoPE of the codebook index), eliminating Moshi's per-level input/output projections; (ii) causal sliding-window attention with fixed-window KV caching, giving constant runtime memory independent of sequence length; and (iii) on-device deployment on the Apple Matrix Coprocessor at ~10 ms per generation step (~16x real time), ~21 MB peak runtime memory, and 329 MB assets. Empirical support includes a component ablation (Table 3), training-curve studies of conditioning, lookahead, window size, and layer ratios (Figs. 2–4), codec-quality metrics on the research configuration (Table 4: UTMOS 3.97, SI-SNR 9.47), an ABX phonetic discriminability check (5.3%), a GPU scaling comparison (Table 2), and a full-system MOS evaluation (+0.28 overall, +0.42 conversational) against the prior production TTS.

Significance. If the claims hold, this is a practically significant contribution: a production-deployed, streaming semantic-to-RVQ detokenizer that runs at ~16x real time with ~21 MB peak runtime memory on the AMX, enabling on-device expressive TTS alongside a sparsely activated foundation model. Strengths worth crediting: a clear architectural delta over Moshi (fully shared depth decoder with DiT stage conditioning, eliminating per-level projections); a controlled, honestly labeled cross-architecture scaling comparison on a common GPU (Table 2); a reasonably complete component ablation (Table 3) plus conditioning, lookahead, window, and layer-ratio sweeps with training curves; and unusually candid disclosure that the research and production configurations differ. That said, the constant-memory property follows largely from construction (fixed-window KV cache), so the genuinely empirical content — quality under the constraint, on the shipped model — is precisely the part not yet demonstrated, which caps the current evidentiary significance.

major comments (4)
  1. [§3.1, Table 1 vs. §3.7, Table 4] The headline claim is a conjunction — constant ~21 MB memory at 16x real time AND synthesis fidelity — but no single model is shown to satisfy both halves. §3.1 states the shipped model 'differs in several hyperparameters (e.g., layer allocation, number of RVQ levels, frame rate, and bitrate)' and that 'unless otherwise stated, the numbers reported refer to the research configuration'; Table 1 (memory/latency) is the production deployment, while Table 4's own note confirms all quality entries (UTMOS 3.97, SI-SNR 9.47) and the 5.3% ABX rate use the research configuration. The production model is compressed along exactly the axes that govern codec fidelity (RVQ levels, bitrate, frame rate). The abstract's claim that the deployed architecture 'maintains synthesis fidelity' is therefore established only by transitivity across two undocumented configs. Please report production-config UTMOS/SI
  2. [§3.8 (Production Human Evaluation)] The +0.28 overall / +0.42 conversational MOS compares the full AFM 3 Core Advanced system — a new 1B-activation foundation model plus the new detokenizer — against the prior TTS. There is no isolated detokenizer A/B, so the gain cannot be attributed to the architecture that is the paper's contribution; it may derive entirely from the improved semantic-token front-end. Either add a same-front-end detokenizer comparison or reframe the MOS result as a system-level outcome. Separately, no listener counts, confidence intervals, or significance testing are given, so the reliability of the 0.1-MOS-resolution claims cannot be judged.
  3. [§3.7 (Audio Quality Evaluation) and Table 4] The text states VisQOL, MOSNet, and SI-SNR are used to measure acoustic quality, but VisQOL and MOSNet results never appear anywhere in the paper. Likewise, the text says ABX error rates are compared 'with three different baseline models' (Mimi decoder, internal transformer decoder, internal AR-GAN), yet only the proposed model's 5.3% is reported. Finally, Table 4 compares 'Ours' at 1.5 kbps / 25 Hz against Moshi/Mimi at 1.1 kbps / 12.5 Hz — not bitrate- or frame-rate-matched — which weakens the 'matching or beating compared codecs' claim in the abstract. Please report the promised numbers and add a bitrate-matched comparison.
  4. [§3.1, Table 2 and surrounding text] (a) The table note and §3.1 text claim 'our model holds generation time, memory, and RTF constant with sequence length,' but the table itself shows generation time scaling 2.40 → 9.64 → 38.59 s — linear, with RTF constant at 0.12. The sentence is contradicted by the data. (b) The flat 1.13 GB memory follows from construction of the fixed-window KV cache, so the scaling table's empirical content is the speed/quality behavior, not the memory constancy. (c) Contribution (4) and the abstract claim superior speed 'including Moshi,' but Moshi does not appear in Table 2, and at 512 tokens the transformer-decoder baseline is faster than the proposed model (RTF 0.04 vs. 0.12) — the speed advantage holds only at long sequence lengths. Please correct the note, add Moshi or drop it from the speed claim, and scope the claim to the long-sequence regime.
minor comments (6)
  1. [Table 3 (Ablation Study)] All rows are evaluated by token-level eval loss only; since Table 4 shows loss and perceptual quality need not track each other, reporting UTMOS or SI-SNR for at least the key ablations (lookahead, unified depth decoder) would strengthen the claim that these components 'validate' audio quality, not just token prediction.
  2. [§3.8] 'Audio cleanliness' (fraction free of tremor, slurs, glitches) is introduced without a definition, annotation protocol, or rater information; 'Navigation reaching 100% clean' is anecdotal as stated.
  3. [§2 / §3] No table specifies the research configuration itself (parameter counts per module, RVQ levels K, codebook size, window W used in the final model). Since the production config is undisclosed, the research config is the only reproducible artifact and should be fully specified.
  4. [§2 (temporal decoder paragraph)] Two empirical claims are asserted without numbers: removing the RVQ autoregressive loop 'leads to a severe drop in generated RVQ token accuracy' (§2), and bidirectional encoding gives 'marginal accuracy improvements' (§3.3). Please quantify or cite a figure.
  5. [Figure 1 and Eqs. (2.1)–(2.7)] Eqs. (2.1)–(2.7): the role of RoPE in Eq. (2.3) — applying rotary encoding of the codebook index to the temporal-decoder output as the sole level identifier — deserves one more sentence, since it is the mechanism distinguishing the model from Moshi. Also clarify whether qt,k in Eq. (2.5) is the codebook embedding of the sampled token rt,k.
  6. [§3.7] Grammar: 'Inspired from Moshi' → 'Inspired by Moshi'; 'in order to measure' can be simplified. The insets in Figures 2–4 are described in captions but their axes are difficult to read at print size.

Circularity Check

0 steps flagged

No constructional circularity: empirical systems paper whose memory, latency, loss, and MOS claims are measured outcomes, not identities forced by definition or self-citation.

full rationale

The paper proposes an engineering architecture (streaming encoder + temporal decoder + single reusable DiT-stage-conditioned depth decoder with fixed-window KV cache) and reports measured latency, peak memory, ablation losses, codec metrics (UTMOS/SI-SNR/ABX), and full-system MOS. Constant memory is an intentional, explicitly stated consequence of fixed-window KV caching (Eqs. 2.1–2.7 and §2), not a disguised fit or a uniqueness theorem imported from the authors; Table 2 simply confirms the designed scaling. Ablations (Table 3) vary components and report eval loss without redefining the target metric as the fit. Quality and MOS comparisons are against external or prior production baselines and are not forced by construction from the paper’s own inputs. Self-references to AFM 3 / Siri Expressive Voices frame deployment context and do not load-bear any uniqueness or derivation step. Config mismatch between research quality numbers and production memory numbers is a validity concern, not circularity. No self-definitional loop, fitted-input-as-prediction, or ansatz-smuggled-via-self-citation is present. Score 0 with empty steps is the correct outcome.

Axiom & Free-Parameter Ledger

6 free parameters · 5 axioms · 1 invented entities

Load-bearing content is architectural and empirical, not axiomatic physics. The claim rests on standard transformer/RVQ practice, several hand-chosen hyperparameters found by ablation, proprietary audio data, and the modeling choice that mean RVQ embeddings plus fixed causal windows suffice for temporal feedback. No new physical entities; the ‘unified depth decoder’ is a design module.

free parameters (6)
  • layer depth ratio (encoder/temporal/depth) = 4-6-2
    Chosen by ablation; 4-6-2 reported as optimal for loss/accuracy/efficiency.
  • temporal lookahead length = 6 tokens (primary); curves for 1–24
    Set by ablation; paper argues 6-token (~240 ms) offset is required unlike Moshi’s single-frame claim.
  • sliding window size W = 128
    Selected empirically; 128 tokens preferred after testing 32–256.
  • learning rate and schedule = 1e-4 cosine
    Hyperparameter search; 1e-4 with cosine annealing and extended peak.
  • RVQ levels K, frame rate, bitrate (research vs production) = research 1.5 kbps @ 25 Hz; production unspecified deltas
    Research table uses 24 kHz, 25 Hz, 1.5 kbps; production changes K/frame rate/bitrate without full disclosure—central device claims depend on these choices.
  • logit scale 1/sqrt(V) = 1/sqrt(V)
    Fixed normalization choice before logits, standard but consequential for training dynamics.
axioms (5)
  • domain assumption Causal sliding-window attention with fixed KV cache preserves sufficient context for high-fidelity autoregressive RVQ synthesis at long durations.
    Invoked throughout §2 and On-Device Deployment; underpins constant-memory claim versus full-context transformers.
  • domain assumption DiT-style stage conditioning (RoPE of codebook index + adaLN) can replace per-level input/output projections without material quality loss.
    Core differentiator vs Moshi; supported by ablation Table 3 and §3.2 but remains an empirical modeling assumption.
  • ad hoc to paper Mean of depth-decoder embeddings is an adequate temporal feedback summary into the next time step (Eqs. 2.6–2.7).
    Specific interface choice of this architecture; removing the AR loop is said to collapse accuracy, but the mean pooling form is a design axiom.
  • standard math Standard transformer components (RMSNorm, SwiGLU, RoPE, GQA, pre-norm, no biases) behave as in LLaMA/Gemma/PaLM literature.
    §3 cites these as drop-in architectural innovations.
  • domain assumption Proprietary internal voice-assistant recordings are a valid training distribution for reported generalization and MOS domains (Siri/Conversational/Navigation).
    Dataset undisclosed by policy; all quality claims rest on it.
invented entities (1)
  • Fully shared reusable depth decoder with DiT stage-k conditioning only no independent evidence
    purpose: Generate all RVQ levels autoregressively without per-level projection parameters.
    Design module rather than a physical entity; independent_evidence is internal ablations and metrics, not an external falsifiable object beyond implementability.

pith-pipeline@v1.2.0-grok45-kimik3 · 15854 in / 4018 out tokens · 79208 ms · 2026-07-30T11:45:24.945881+00:00 · methodology

0 comments
read the original abstract

Siri Expressive Voices synthesize rich, configurable speech in real time and entirely on device, powered by AFM 3 Core Advanced, Apple's most powerful on-device foundation model. This work presents the memory-efficient audio synthesis architecture behind that capability: a detokenizer that converts the semantic audio tokens emitted by the foundation model into high-fidelity audio within the tight compute and memory budget of the Apple Matrix Coprocessor (AMX). We convert semantic audio tokens to a residual vector quantization (RVQ) representation with a three-component design, a streaming encoder, a temporal decoder, and a depth decoder, that systematically decouples temporal and depth processing. A single reusable depth decoder with Diffusion Transformer (DiT)-style stage conditioning generates all RVQ levels autoregressively, replacing the dedicated per-level decoders of prior multi-decoder architectures, while causal sliding window attention with fixed-window key-value caching yields constant memory complexity independent of sequence length. Deployed on the AMX, the detokenizer sustains roughly 10 ms per generation step, about 16x faster than real time, with a peak runtime memory of only 21 MB and 329 MB of on-device assets, enabling continuous streaming synthesis of 20-320 seconds of audio. This constant, small footprint replaces the linear and quadratic memory scaling of conventional transformer- and GAN-based approaches. Ablation studies validate the key architectural components, and audio quality assessment confirms that the architecture maintains synthesis fidelity while achieving efficiency gains over existing methods. Operating at a 1-billion-parameter activation size within AFM 3 Core Advanced, it improves Mean Opinion Score by +0.28 overall (4.15 vs. 3.87) and by +0.42 on conversational speech (4.24 vs. 3.82) over the prior on-device text-to-speech system.

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

20 extracted references · 14 linked inside Pith

  1. [1]

    Musiclm: Generating music from text.arXiv preprint arXiv:2301.11325,

    Andrea Agostinelli, Timo I Denk, Zalán Borsos, Jesse Engel, Mauro Verzetti, Antoine Caillon, Qingqing Huang, Aren Jansen, Adam Roberts, Marco Tagliasacchi, et al. Musiclm: Generating music from text.arXiv preprint arXiv:2301.11325,

    Pith/arXiv arXiv

  2. [8]

    com/deepmind-media/gemma/gemma-2-report.pdf

    URLhttps://storage.googleapis. com/deepmind-media/gemma/gemma-2-report.pdf. Technical Report. Andrew Hines, Jan Skoglund, Andrew Kokaram, and N. Harte. Visqol: an objective speech quality model.EURASIP Journal on Audio, Speech, and Music Processing, 2015:1–13,

    2015

  3. [11]

    Kushal Lakhotia, Evgeny Kharitonov, Wei-Ning Hsu, Yossi Adi, Adam Polyak, Benjamin Bolte, Tu-Anh Nguyen, Jade Copet, Alexei Baevski, Adelrahman Mohamed, and Emmanuel Dupoux

    URLhttps://arxiv.org/abs/2209.15352. Kushal Lakhotia, Evgeny Kharitonov, Wei-Ning Hsu, Yossi Adi, Adam Polyak, Benjamin Bolte, Tu-Anh Nguyen, Jade Copet, Alexei Baevski, Adelrahman Mohamed, and Emmanuel Dupoux. Generative spoken language modeling from raw audio,

    Pith/arXiv arXiv

  4. [15]

    AaronvandenOord, SanderDieleman, HeigaZen, KarenSimonyan, OriolVinyals, AlexGraves, NalKalchbrenner, An- drewSenior, andKorayKavukcuoglu

    URLhttps://arxiv.org/abs/2110.10139. AaronvandenOord, SanderDieleman, HeigaZen, KarenSimonyan, OriolVinyals, AlexGraves, NalKalchbrenner, An- drewSenior, andKorayKavukcuoglu. Wavenet: Agenerativemodelforrawaudio.arXiv preprint arXiv:1609.03499,

    Pith/arXiv arXiv

  5. [16]

    Librispeech: An asr corpus based on public domain audio books.2015 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 5206–5210,

    Vassil Panayotov, Guoguo Chen, Daniel Povey, and Sanjeev Khudanpur. Librispeech: An asr corpus based on public domain audio books.2015 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 5206–5210,

    2015

  6. [18]

    Eval- uating speech features with the minimal-pair abx task: analysis of the classical mfc/plp pipeline

    Thomas Schatz, Vijayaditya Peddinti, Francis Bach, Aren Jansen, Hynek Hermansky, and Emmanuel Dupoux. Eval- uating speech features with the minimal-pair abx task: analysis of the classical mfc/plp pipeline. InInterspeech 2013, pages 1781–1785,

    2013

  7. [20]

    Haibin Wu, Naoyuki Kanda, Sefik Emre Eskimez, and Jinyu Li

    URLhttps://arxiv.org/abs/2002.05202. Haibin Wu, Naoyuki Kanda, Sefik Emre Eskimez, and Jinyu Li. Ts3-codec: Transformer-based simple streaming single codec,

    Pith/arXiv arXiv 2002

  8. [21]

    Haibin Wu, Bach Viet Do, Naveen Suda, Julian Chan, C R Madhavan, Gene-Ping Yang, Yi-Chiao Wu, Naoyuki Kanda, Yossef Adi, Xin Lei, Yue Liu, Florian Metze, and Yuzong Liu

    URLhttps://arxiv.org/abs/2411.18803. Haibin Wu, Bach Viet Do, Naveen Suda, Julian Chan, C R Madhavan, Gene-Ping Yang, Yi-Chiao Wu, Naoyuki Kanda, Yossef Adi, Xin Lei, Yue Liu, Florian Metze, and Yuzong Liu. T-mimi: A transformer-based mimi decoder for real-time on-phone tts,

    Pith/arXiv arXiv

  9. [22]

    Biao Zhang and Rico Sennrich

    URLhttps://arxiv.org/abs/2601.20094. Biao Zhang and Rico Sennrich. Root mean square layer normalization.CoRR, abs/1910.07467,

    arXiv 1910

  10. [23]

    AppleandtheApplelogoaretrademarksofAppleInc.,registeredintheU.S.andothercountriesandregions

    URLhttp: //arxiv.org/abs/1910.07467. AppleandtheApplelogoaretrademarksofAppleInc.,registeredintheU.S.andothercountriesandregions. 11

    Pith/arXiv arXiv 1910

  11. [2015]

    William Peebles and Saining Xie

    URLhttps://api.semanticscholar.org/CorpusID:2191379. William Peebles and Saining Xie. Scalable diffusion models with transformers. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 4195–4205, 2023a. William Peebles and Saining Xie. Scalable diffusion models with transformers, 2023b. URLhttps://arxiv.org/abs/2212. 09748. Ryan...

    2019

  12. [2018]

    Diffwave: A versatile diffusion model for audio synthesis.arXiv preprint arXiv:2009.09761,

    Zhifeng Kong, Wei Ping, Jiaming Huang, Kexin Zhao, and Bryan Catanzaro. Diffwave: A versatile diffusion model for audio synthesis.arXiv preprint arXiv:2009.09761,

    Pith/arXiv arXiv 2009

  13. [2019]

    URL https://arxiv.org/abs/1904.08352. Meta. Llama 3: The next generation of open models. Technical report, Meta,

    Pith/arXiv arXiv 1904

  14. [2020]

    Zalán Borsos, RaphaëlMarinier, DamienVincent, Eugene Kharitonov, Olivier Pietquin, Matt Sharifi, DominikRoblek, Olivier Teboul, David Grangier, Marco Tagliasacchi, et al

    URLhttps://arxiv.org/abs/2004.05150. Zalán Borsos, RaphaëlMarinier, DamienVincent, Eugene Kharitonov, Olivier Pietquin, Matt Sharifi, DominikRoblek, Olivier Teboul, David Grangier, Marco Tagliasacchi, et al. Audiolm: a language modeling approach to audio generation.IEEE/ACM Transactions on Audio, Speech, and Language Processing, 31:2523–2533,

    Pith/arXiv arXiv 2004

  15. [2021]

    Doyup Lee, Chiheon Kim, Saehoon Kim, Minsu Cho, and Wook-Shin Han

    URLhttps://arxiv.org/abs/2102.01192. Doyup Lee, Chiheon Kim, Saehoon Kim, Minsu Cho, and Wook-Shin Han. Autoregressive image generation using residual quantization.Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11523–11532,

    Pith/arXiv arXiv

  16. [2022]

    Zach Evans, Julian Parker, CJ Carr, Zack Zukowski, Josiah Taylor, and Jordi Pons

    URL https://arxiv.org/abs/2210.13438. Zach Evans, Julian Parker, CJ Carr, Zack Zukowski, Josiah Taylor, and Jordi Pons. Long-form music generation with latent diffusion.arXiv preprint arXiv:2404.10301,

    Pith/arXiv arXiv

  17. [2023]

    Moshi: a speech-text foundation model for real-time dialogue.arXiv preprint arXiv:2410.00037,

    Alexandre Défossez, Yossi Adi, Jade Copet, Gabriel Synnaeve, Itai Elbaz, Tal Remez, Olivier Teboul, David Grangier, Artyom Cattan, Felix Maurer, et al. Moshi: a speech-text foundation model for real-time dialogue.arXiv preprint arXiv:2410.00037,

    Pith/arXiv arXiv

  18. [2024]

    Jukebox: A generative model for music.arXiv preprint arXiv:2005.00341,

    Prafulla Dhariwal, Heewoo Jun, Christine Payne, Jong Wook Kim, Alec Radford, and Ilya Sutskever. Jukebox: A generative model for music.arXiv preprint arXiv:2005.00341,

    Pith/arXiv arXiv 2005

  19. [2025]

    Nal Kalchbrenner, Erich Elsen, Karen Simonyan, Seb Noury, Norman Norm, Eamonn Lockhart, Florian Stimberg, Aaron van den Oord, Sander Dieleman, and Koray Kavukcuoglu

    URLhttps://arxiv.org/abs/2501.02086. Nal Kalchbrenner, Erich Elsen, Karen Simonyan, Seb Noury, Norman Norm, Eamonn Lockhart, Florian Stimberg, Aaron van den Oord, Sander Dieleman, and Koray Kavukcuoglu. Efficient neural audio synthesis.arXiv preprint arXiv:1802.08435,

    Pith/arXiv arXiv

  20. [2026]

    URLhttps://arxiv.org/abs/2603.08823. Google. PaLM: Scaling Language Modeling with Pathways,

    arXiv