Pith. sign in

REVIEW 4 major objections 6 minor 24 references

WhisperKit: On-device Real-time ASR with Billion-Scale Transformers

T0 review · 4 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read WhisperKit claims a billion-parameter Whisper model can run entirely on-device at 0.46 s latency and 2.2% WER, matching or beating cloud-hosted ASR systems.

desk verdict Solid engineering paper with a real on-device streaming Whisper system, but the headline 'significantly outperforms cloud systems' is undercut by the paper's own Table 2 and by missing variance and eval-set details. read the letter →

arxiv 2507.10860 v1 pith:RL26NM6T submitted 2025-07-14 cs.SD cs.CLeess.AS

Atila Orhon , Arda Okan , Berkin Durmus , Zach Nagengast , Eduardo Pacheco This is my paper
classification cs.SDcs.CLeess.AS
keywords automaticspeechrecognitionon-deviceinferencestreamingASRWhisperAppleNeuralEnginemodelcompressionpalettizationspeculativedecoding
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

The paper tries to establish that a billion-parameter speech recognition model can run entirely on a laptop's neural accelerator and still beat mainstream cloud transcription services on latency and accuracy, the two metrics that decide commercial deployment. The authors build WhisperKit on Whisper Large v3 Turbo, restructure the audio encoder's attention so it streams instead of chewing on 30-second windows, and compress the weights from 1.6 GB to 0.6 GB with a technique called OD-MBP, outlier-decomposed mixed-bit palettization, that keeps word error rate within about 1% of the uncompressed model. In their benchmark against a frontier cloud API, a proprietary ASR API, and a cloud-hosted copy of the same Whisper model, WhisperKit reports the lowest latency, 0.46 s mean per-word hypothesis latency, and the highest confirmed-text accuracy, 2.2% WER. If those measurements hold, on-device ASR becomes commercially competitive with cloud ASR for live captioning, dictation, meetings, and medical scribes, with the added privacy and reliability of keeping audio on the device.

What carries the argument

Three mechanisms carry the argument. First, a block-diagonal self-attention mask, called d750, is applied to the audio encoder during self-distillation on Common Voice: attention is causal inside 15-second chunks, and the encoder output for a completely zero-padded 15-second block is computed once at compile time and reused, cutting audio encoder latency from roughly 602 ms to 218 ms on the M3 Max Neural Engine with WER within 1% of the unmasked model. Second, the LocalAgreement streaming policy advances the transcript cursor only on tokens confirmed across consecutive hypotheses, yielding two output streams: stable confirmed text and low-latency hypothesis text. Third, OD-MBP splits each weight tensor into a dense inlier block stored as a low-bit lookup table plus a sparse float16 outlier block, so quantization error is concentrated in a small fraction of weights; this shrinks the model from 1.6 GB to 0.6 GB while keeping WER degradation within about one percentage point. The system also uses stateful key-value caches that update in place on the Neural Engine, reducing text decoder forward-pass latency by 45% and energy by 75%.

What would settle it

The decisive experiment is a controlled re-run of the latency comparison using a third-party harness that records timestamps from a shared clock, randomizes trial order, and adds calibrated network jitter to the cloud APIs; if the reported 0.46 s mean latency and 2.2% WER advantage over the cloud systems do not reproduce under that protocol, the headline comparison is a measurement artifact. A second check is running WhisperKit on older supported Neural Engine devices and on the full 100-language Common Voice set, since a WER increase beyond the claimed 1% or latency above 0.46 s would falsify the generality of the results.

Watch

Extended reading notes

Core claim

On its own terms, the paper claims that a task-focused billion-parameter model can outrun frontier and proprietary cloud ASR on the metrics that matter for deployment. The evidence is WhisperKit: the audio encoder is self-distilled with a block-diagonal causal attention mask so it can stream over partial audio while caching zero-padded blocks as silence; the text decoder uses the LocalAgreement policy to publish stable confirmed text alongside low-latency hypothesis text; and the weights are compressed with outlier-decomposed mixed-bit palettization from 1.6 GB to 0.6 GB with less than 1% WER change. Benchmarked on an M3 Max MacBook Pro against one frontier cloud API, one proprietary ASR API, and a cloud-hosted version of the same Whisper model, WhisperKit reports the lowest latency, 0.46 s mean per-word hypothesis latency, and the lowest confirmed-text word error rate, 2.2%.

Load-bearing premise

The load-bearing premise is that the benchmark is a fair and representative comparison: cloud systems were measured over uncontrolled network conditions while WhisperKit ran locally on one M3 Max MacBook Pro, and accuracy was evaluated on datasets and 22 languages where the base model already performs best.

Editorial extensions

If this is right

  • Real-time ASR for live captions and dictation no longer needs a network round trip: a 0.6 GB model on a laptop delivers 0.46 s hypothesis latency, competitive with or better than the cloud APIs tested.
  • Streaming Whisper-style encoders can reuse a cached silent 15-second block at compile time, cutting audio encoder latency by about 65% while staying within 1% WER of the original model.
  • OD-MBP compression places a billion-parameter transformer under the 2 GB memory threshold that the paper argues is required for broad on-device support, reducing the model file from 1.6 GB to 0.6 GB.
  • Stateful key-value cache handling cuts decoder forward-pass latency by 45% and energy per forward pass by 75%, addressing the battery and thermal constraints of on-device inference.
  • The LocalAgreement policy yields both a stable confirmed text stream and a low-latency hypothesis stream, and the paper reports the hypothesis stream needs far fewer corrections for WhisperKit than for a cloud-hosted Whisper service.

Reading between the lines

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

  • The paper does not test whether its latency and accuracy figures generalize beyond one M3 Max MacBook Pro; profiling on the oldest supported Apple Neural Engine devices and on other hardware would show whether the 0.46 s and 2.2% numbers are tied to this specific chip.
  • Because the cloud APIs were measured over uncontrolled network conditions, the local model's 0.46 s figure is likely more consistent than any cloud mean; a controlled network-emulation study would quantify that consistency advantage, which the paper only gestures at.
  • OD-MBP is presented as a generic weight-compression format but is demonstrated only on Whisper; evaluating it on non-ASR transformers would show whether the under-1% accuracy retention transfers to other tasks.
  • The d750 mask uses 15-second blocks as a fixed latency-accuracy tradeoff; the paper's own results show shorter blocks cut latency further but degrade accuracy sharply, so applications could re-tune the block size for their target operating point.
Share X Bluesky LinkedIn Reddit HN

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. The paper presents WhisperKit, an optimized on-device ASR system for Apple Neural Engine based on Whisper Large v3 Turbo. The contributions are: (i) a streaming audio encoder obtained by self-distillation with block-diagonal self-attention masks (d750), which enables silence caching and reduces encoder latency; (ii) a streaming text decoder using the LocalAgreement policy and stateful KV-cache execution; (iii) a compression scheme called OD-MBP that decomposes weights into palettized inliers and sparse float16 outliers, reducing the model from 1.6 GB to 0.6 GB; and (iv) a benchmark against cloud APIs (OpenAI gpt-4o-transcribe, Deepgram nova-3, Fireworks large-v3-turbo) reporting that WhisperKit matches the lowest latency (0.46 s) and achieves the highest accuracy (2.2% WER). The paper also reports speculative-decoding speedups, energy measurements, and multilingual results on a subset of Common Voice.

Significance. If the headline claims are correct, WhisperKit would demonstrate that a 1-billion-parameter on-device ASR system can match or beat leading cloud-based systems in both latency and accuracy, with important implications for privacy, cost, and energy consumption. The engineering artifacts are concrete and potentially reusable: the ANE reimplementation, the block-diagonal masking with silence caching, the stateful KV-cache optimization, and the OD-MBP compression are all described in enough detail to be adapted, and the code is released. However, the central empirical claims are currently under-supported. The compression claim is contradicted by the paper's own Table 2, the latency methodology relies on oracle timestamps on a single clean corpus and mixes different stream types, and the accuracy evaluation does not name its corpus or provide error bars. These issues are load-bearing for the paper's headline, so the manuscript needs major revision.

major comments (4)
  1. [2.2.2 (Table 2)] The text states that 'OD-MBP, even when combined with the d750 optimization, retains WER within 1% of the original uncompressed model.' Table 2 contradicts this: Common Voice 17 English moves from 12.13 to 14.21 (+2.08) and earnings22 moves from 11.55 to 12.72 (+1.17). These are not within 1%. The claim must be revised (for example, to 'within 2%' or 'on librispeech and earnings22') or the table/experiment must be corrected. As written, the compression contribution's headline property is not established.
  2. [3.1 (Figures 6 and 7)] The latency comparison uses TIMIT ground-truth word timestamps to set the transcript cursor because Fireworks and OpenAI do not return predicted timestamps. This means the measured latency partly reflects how well each system tracks oracle word boundaries rather than end-to-end user-perceived delay, and it compares incompatible stream types: hypothesis text for WhisperKit, Deepgram, and Fireworks, but confirmed text for OpenAI. Additionally, TIMIT is a clean, read-speech, 8 kHz corpus of short utterances; the headline '0.46s' is a point estimate on this corpus with no repeated trials, confidence intervals, or per-utterance distributions. Please state these limitations explicitly and provide variance information.
  3. [3.2 (Figure 8)] The confirmed-text WER values in Figure 8 are reported without stating which corpus produced them. The abstract claims '2.2% WER' while the Section 3.2 text says '2% WER.' Without naming the evaluation corpus and the number of utterances, the central accuracy claim is not reproducible. If the corpus is TIMIT, the claim is limited in scope and should be described as such; if it is a different corpus, it must be identified and the abstract/text numbers reconciled.
  4. [2.1.1 (Figure 2)] The multilingual evaluation is restricted to the 20 languages where the original model achieves the lowest WER/CER plus Japanese and Chinese. This subset is selected after inspecting the original model's scores on the same Common Voice 17 test split that is used for the final reporting, so the reported multilingual retention is not representative of the model's 100-language capability. The self-distillation is also fine-tuned on only 5 languages. Please report full language-wise results or a pre-specified subset, and note that the presented numbers are conditional on the selection criterion.
minor comments (6)
  1. [Abstract / 3.1] The latency is reported as 0.46s in the abstract and 0.45 seconds in Section 3.1 and Figure 6; please reconcile these values.
  2. [Abstract / 3.2] The accuracy is reported as 2.2% WER in the abstract and 2% WER in Section 3.2; please reconcile these values.
  3. [Table 1 caption] The caption contains a typo: 'c250 latency is multipled by 6' should read 'multiplied by 6.'
  4. [Section 2] The sentence ending 'on Apple devices. 1.' contains a stray '1.' that appears to be a numbering artifact; please remove it.
  5. [References] The earnings22 reference is incomplete in the reference list; provide the full citation with venue and year.
  6. [Section 3.2] The sentence 'WhisperKit latency is measured on a MacBook Pro with M3 Max chip on the Neural Engine' appears in the accuracy section; please clarify whether this hardware statement applies to the accuracy measurements, the latency measurements, or both.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: WhisperKit's headline accuracy and latency claims are measured against external cloud APIs, not derived from fitted parameters or self-citations.

full rationale

The paper's central claims are empirical benchmark results, not outputs of a derivation whose inputs contain those claims. Table 1 and Table 2 report measured WER and latency for attention-mask and compression variants; Figures 6-8 compare WhisperKit against external cloud APIs. The d750 block size, OD-MBP outlier threshold, and the 22-language subset are engineering/tuning choices evaluated on the same benchmarks that produce the reported numbers, so the results describe a tuned system rather than an independent prediction; however, none of these choices forces the reported WER or latency by construction. The self-citations (ane-transformers, MBP) point to prior implementation artifacts and are not load-bearing: WhisperKit's performance is independently measured, and the authors state that their latency results approximately match the cloud providers' officially reported latencies. No equation sets a claimed result equal to a fitted parameter or to a cited result, so no circular step is exhibited. The TIMIT-based latency methodology and the Table 2 inconsistency (+2.08 WER on Common Voice 17 English for OD-MBP d750, contradicting the 'within 1%' text) are benchmark-validity/correctness concerns, not circularity.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

The central claims are empirical system measurements rather than a mathematical derivation. The ledger lists the hand-chosen design parameters (block size, outlier threshold, bit widths, language subsets) and the domain assumptions about ANE behavior, silence caching, timestamp alignment, and benchmark stability that the results depend on.

free parameters (5)
  • Attention block size d (d750) = 750 (15-second audio blocks)
    Chosen by sweeping block sizes in Table 1; d250 and d500 degrade WER substantially, and d750 is selected as the latency-accuracy tradeoff.
  • OD-MBP outlier threshold = 3 standard deviations from the mean
    Outliers are defined as weights more than 3 sigma from the mean; this threshold determines the sparse outlier fraction and is not justified by theory.
  • OD-MBP palettization bit widths = not reported
    The inlier branch uses low-bit palettization and the outlier branch uses fp16, but the exact bit widths are not specified even though they determine the 0.6 GB file size and WER impact.
  • Self-distillation language subset = English, German, Japanese, Chinese, French
    The d750 encoder is fine-tuned only on these five languages from Common Voice 17; the choice affects the multilingual results in Figure 2.
  • Benchmark language subset for multilingual accuracy = 20 lowest-WER languages plus Japanese and Chinese
    The authors explicitly restrict the multilingual evaluation to languages where the original model already performs best, which makes the multilingual comparison more favorable than an evaluation over all 100 claimed languages.
assumptions (5)
  • domain assumption Core ML and the Apple Neural Engine calculate attention and KV cache updates in a numerically correct way for the stateful model implementation.
    Section 2.2.1 relies on Apple's Stateful Models feature for the 45% decoder latency reduction and 75% energy reduction, without proving numerical equivalence to a reference implementation.
  • domain assumption Zero-padded audio blocks produce encoder outputs that can be precomputed at compile time and reused during streaming.
    Section 2.1.1 introduces silence caching, which assumes that a block-diagonal mask makes a silent block's encoder output independent of the streaming context.
  • domain assumption TIMIT ground-truth word timestamps provide a fair word-cursor alignment for all cloud APIs under comparison.
    Section 3.1 replaces predicted timestamps with ground truth because Fireworks and OpenAI do not return timestamps; this may favor systems whose internal timestamp handling differs.
  • domain assumption Self-distillation with a masked attention encoder preserves the teacher model's knowledge for the unmasked language distribution.
    Section 2.1.1 assumes that fine-tuning on five languages with d750 masking is a scalable way to improve multilingual accuracy without losing English or other-language performance.
  • domain assumption Cloud API latency measurements are stable enough for point-wise comparison with a single local device measurement.
    Section 3.1 reports mean latencies without confidence intervals, significance tests, or control for network conditions on the cloud API side.

how reviews work

0 comments
Cite this review

Pith. "Pith review of WhisperKit: On-device Real-time ASR with Billion-Scale Transformers." pith.science (2026). https://pith.science/paper/RL26NM6T

@misc{pith2026250710860,
  author       = {Pith},
  title        = {Pith review of: WhisperKit: On-device Real-time ASR with Billion-Scale Transformers},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/RL26NM6T}},
  note         = {Machine review of arXiv:2507.10860}
}
read the original abstract

Real-time Automatic Speech Recognition (ASR) is a fundamental building block for many commercial applications of ML, including live captioning, dictation, meeting transcriptions, and medical scribes. Accuracy and latency are the most important factors when companies select a system to deploy. We present WhisperKit, an optimized on-device inference system for real-time ASR that significantly outperforms leading cloud-based systems. We benchmark against server-side systems that deploy a diverse set of models, including a frontier model (OpenAI gpt-4o-transcribe), a proprietary model (Deepgram nova-3), and an open-source model (Fireworks large-v3-turbo).Our results show that WhisperKit matches the lowest latency at 0.46s while achieving the highest accuracy 2.2% WER. The optimizations behind the WhisperKit system are described in detail in this paper.

Figures

Figures reproduced from arXiv: 2507.10860 by the authors.

Figure 1
Figure 1. Block-diagonal (d*) and block-causal (c*) masks for Whisper Audio Encoder 1500x1500 (30 seconds) Self-attention Matrix. Mask TFLOPs Audio Encoder Latency (ms) librispeech-test.clean (WER) earnings222 (WER) original 2.27 612 1.93 11.55 c250 2.34 117 ∗ 6 = 702 2.32 (+0.39) 12.89 (+1.34) d750 1.04 218 2.25 (+0.32) 12.85 (+1.30) d500 0.68 125 3.39 (+1.46) 13.86 (+2.31) d250 0.33 50 25.58 (+23.65) 39.75 (+28.20) [PITH_F… view at source ↗
Figure 2
Figure 2. Multilingual WER (Word Error Rate) and CER (Character Error Rate) before (original) and after (d750) self￾distillation of Whisper Large v3 Turbo. Metrics are reported for the 20 languages with the lowest WER/CER in the original model, plus Japanese and Chinese, on a subset of the Common Voice 17 test set. 4 . Languages marked with † were used during self-distillation. man) when compared to the original model for lan… view at source ↗
Figure 3
Figure 3. Speedup factors for Whisper Text Decoder with Spec￾ulative Decoding. Theoretical speedup (tokens/step) is the token acceptance rate of the verification step and signifies the speedup for an idealized hardware with infinite compute and limited memory bandwidth. The practical slowdowns are due to the verification overhead (processing multiple tokens in parallel) and drafter over￾head (forward pass latency of the draft… view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Compile time with OD-MBP. The original weight ten￾sor W is decomposed into a dense inlier block Winlier (green) and a sparse outlier block Woutlier (red). Outliers are defined as weight values that are more than 3 standard deviations away from the mean. Inliers are pal…
Figure 5
Figure 5. Figure 5: Inference with OD-MBP. The float16 input activa￾tion X is dispatched to two parallel operators: (i) Dense path. A Linear layer retrieves the low-bit palettized inlier weights Q(Winlier), dequantizes them on the fly, and performs a dense matrix–vector product in float16…
Figure 6
Figure 6. Figure 6: shows the per-word latency for the hypothesis text stream. WhisperKit and Fireworks are the fastest, achieving a roughly equal mean latency of 0.45 seconds. Deepgram achieves 0.83 seconds and becomes the third fastest sys￾tem. Note that OpenAI does not support hypothes…
Figure 7
Figure 7. Figure 7: Per-word Latency Histogram for Confirmed Text. When hypothesis text results are not considered, all benchmarked systems achieve a similarly high latency of 1.7 seconds. Fireworks API does not support partial resul confirmations for any result and is excluded from this …

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

24 extracted references · 15 canonical work pages

  1. [1]

    Core ml, 2017

    Apple, I. Core ml, 2017. URL https://developer.apple.com/documentation/coreml

    work page 2017

  2. [2]

    Use core ml tools for machine learning model compression, 2023

    Apple, I. Use core ml tools for machine learning model compression, 2023. URL https://developer.apple.com/videos/play/wwdc2023/10047/

    work page 2023

  3. [3]

    Apple intelligence foundation language models, 2024 a

    Apple, I. Apple intelligence foundation language models, 2024 a . URL https://arxiv.org/abs/2407.21075

    arXiv 2024

  4. [4]

    Stateful models, 2024 b

    Apple, I. Stateful models, 2024 b . URL https://apple.github.io/coremltools/docs-guides/source/stateful-models.html

    work page 2024

  5. [5]

    M., and Weber, G

    Ardila, R., Branson, M., Davis, K., Henretty, M., Kohler, M., Meyer, J., Morais, R., Saunders, L., Tyers, F. M., and Weber, G. Common voice: A massively-multilingual speech corpus. In Proceedings of the 12th Conference on Language Resources and Evaluation (LREC 2020), pp.\ 4211--4215, 2020

    work page 2020

  6. [6]

    Recurrent drafter for fast speculative decoding in large language models, 2024

    Cheng, Y., Zhang, A., Zhang, X., Wang, C., and Wang, Y. Recurrent drafter for fast speculative decoding in large language models, 2024. URL https://arxiv.org/abs/2403.09919

    arXiv 2024

  7. [7]

    and Orhon, A

    Cuenca, P. and Orhon, A. Stable diffusion xl on mac with advanced core ml quantization, 2023. URL https://huggingface.co/blog/stable-diffusion-xl-coreml#

    work page 2023

  8. [8]

    Deepgram latency measurement, 2025 a

    Deepgram. Deepgram latency measurement, 2025 a . URL https://developers.deepgram.com/docs/measuring-streaming-latency

    work page 2025

Show all 24 references
  1. [9]

    Nova 3, 2025 b

    Deepgram. Nova 3, 2025 b . URL https://deepgram.com/learn/introducing-nova-3-speech-to-text-api

    2025

  2. [10]

    Llm.int8(): 8-bit matrix multiplication for transformers at scale, 2022

    Dettmers, T., Lewis, M., Belkada, Y., and Zettlemoyer, L. Llm.int8(): 8-bit matrix multiplication for transformers at scale, 2022. URL https://arxiv.org/abs/2208.07339

    2022 arXiv

  3. [11]

    Fireworks streaming latency, 2025 a

    Fireworks. Fireworks streaming latency, 2025 a . URL https://fireworks.ai/blog/streaming-audio-launch

    2025

  4. [12]

    Whisper-large-v3-turbo, 2025 b

    Fireworks. Whisper-large-v3-turbo, 2025 b . URL https://fireworks.ai/models/fireworks/whisper-v3-turbo

    2025

  5. [13]

    , Lamel, Lori F

    Garofolo, John S. , Lamel, Lori F. , Fisher, William M. , Fiscus, Jonathan G. , Pallett, David S. , and Dahlgren, Nancy L. Timit acoustic-phonetic continuous speech corpus. https://catalog.ldc.upenn.edu/LDC93S1, 1993. Linguistic Data Consortium, LDC93S1

    1993

  6. [14]

    Moonshine: Speech recognition for live transcription and voice commands, 2024

    Jeffries, N., King, E., Kudlur, M., Nicholson, G., Wang, J., and Warden, P. Moonshine: Speech recognition for live transcription and voice commands, 2024. URL https://arxiv.org/abs/2410.15608

    2024 arXiv

  7. [15]

    Low-latency sequence-to-sequence speech recognition and translation by partial hypothesis selection, 2020

    Liu, D., Spanakis, G., and Niehues, J. Low-latency sequence-to-sequence speech recognition and translation by partial hypothesis selection, 2020. URL https://arxiv.org/abs/2005.11185

    2020 arXiv

  8. [16]

    Turning whisper into real-time transcription system, 2023

    Macháček, D., Dabre, R., and Bojar, O. Turning whisper into real-time transcription system, 2023. URL https://arxiv.org/abs/2307.14743

    2023 arXiv

  9. [17]

    Gpt-4o transcribe, 2025

    OpenAI. Gpt-4o transcribe, 2025. URL https://platform.openai.com/docs/models/gpt-4o-transcribe

    2025

  10. [18]

    Deploying transformers on the apple neural engine, 2021

    Orhon, A., Joergensen, M., Lillethorup, M., Vestergaard, J., and Jagadeesh, V. Deploying transformers on the apple neural engine, 2021. URL https://machinelearning.apple.com/research/panoptic-segmentation

    2021

  11. [19]

    Apple neural engine transformers, 2022

    Orhon, A., Wadhwa, A., Kim, Y., Rossi, F., and Jagadeesh, V. Apple neural engine transformers, 2022. URL https://machinelearning.apple.com/research/neural-engine-transformers

    2022

  12. [20]

    W., Xu, T., Brockman, G., McLeavey, C., and Sutskever, I

    Radford, A., Kim, J. W., Xu, T., Brockman, G., McLeavey, C., and Sutskever, I. Robust speech recognition via large-scale weak supervision, 2022. URL https://arxiv.org/abs/2212.04356

    2022 arXiv

  13. [21]

    earnings-22: A practical benchmark for accents in the wild

    Rio, M. D., Ha, P., McNamara, Q., Miller, C., and Chandra, S. "earnings-22: A practical benchmark for accents in the wild", 2022

    2022

  14. [22]

    Bitsfusion: 1.99 bits weight quantization of diffusion model, 2024

    Sui, Y., Li, Y., Kag, A., Idelbayev, Y., Cao, J., Hu, J., Sagar, D., Yuan, B., Tulyakov, S., and Ren, J. Bitsfusion: 1.99 bits weight quantization of diffusion model, 2024. URL https://arxiv.org/abs/2406.04333

    2024 arXiv

  15. [23]

    Conformer-based speech recognition on extreme edge-computing devices

    Xu, M., Jin, A., Wang, S., Su, M., Ng, T., Mason, H., Han, M., Lei, Z., Deng, Y., Huang, Z., and Krishnamoorthy, M. Conformer-based speech recognition on extreme edge-computing devices. In NAACL, 2024. URL https://arxiv.org/abs/2312.10359

    2024 arXiv

  16. [24]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.