Pith. sign in

REVIEW 3 major objections 5 minor 34 references

Pushing the Limits of Beam Search Decoding for Transducer-based ASR models

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read This paper claims that fully batched, trie-based beam search makes Transducer beam decoding nearly as fast as greedy decoding, with 14–30% relative WER gains from n-gram shallow fusion.

desk verdict A solid engineering paper with real speedups; the core claims hold up, but the hash-collision parameters and blank-scoring normalization need a closer look before I'd call the universality claim fully supported. read the letter →

arxiv 2506.00185 v1 pith:7I7JBVHI submitted 2025-05-30 eess.AS cs.AIcs.CLcs.LGcs.SD

Lilit Grigoryan , Vladimir Bataev , Andrei Andrusenko , Hainan Xu , Vitaly Lavrukhin , Boris Ginsburg This is my paper · ORCID
classification eess.AScs.AIcs.CLcs.LGcs.SD
keywords automaticspeechrecognitiontransducermodelsbeamsearchdecodingGPU-acceleratedinferencehypothesistrieincrementalhashingshallowfusioncontextbiasing
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

Beam search for Transducer-based speech recognition has traditionally been five to ten times slower than greedy decoding, which has kept it out of production streaming systems. This paper claims to close almost all of that gap by making beam search fully batched: hypotheses live in a trie-like tensor structure, equality between transcripts is checked with an incremental hash, every expansion is processed in parallel, and the whole decode loop is replayed with CUDA graphs. The resulting ALSD++ and AES++ decoders narrow the whole-system speed gap to 10–20% while improving WER by 14–30% relative to greedy when combined with an external n-gram language model. The same machinery also gives a new blank-scoring rule for shallow fusion that adds up to 11% relative WER improvement over existing fusion in low-resource settings and transfers to Token-and-Duration Transducer models. A sympathetic reader would take the paper's thesis to be that beam search can be an inexpensive accuracy lever again, not a bottleneck.

What carries the argument

The load-bearing object is a trie-like hypothesis store the paper calls BatchedBeamHyps: it keeps all transcripts in two 3D tensors (transcripts and transcripts ptrs) so that identical prefixes are shared and full strings are reconstructed by following back-pointers. Equality of hypotheses is checked by an incremental polynomial hash $H_{t+1}=(H_t\cdot P+T_{t+1}) \bmod M$, with length and latest-token checks as a guard, turning transcript comparison from linear to constant time. Around this, the method batches all expansions and pruning operations for the whole beam and uses CUDA Graphs to replay the fixed decode sequence. The other load-bearing piece is a blank-scoring rule for shallow fusion: instead of leaving the blank score to the ASR alone, it scales LM probabilities by $1-p[\varnothing]$ and the blank's log-probability by $1+\lambda$, so a heavy LM weight does not push the decoder toward deleting words.

What would settle it

Run ALSD++ with a second independent hash modulus (or full transcript comparison) on a large decoded corpus with a large beam and a skewed token distribution, and count how often two distinct transcripts with the same length and latest token collide; if the rate exceeds about 0.1% or degrades WER relative to exact merging, the constant-time merging claim needs a collision-resolution strategy.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is a universal acceleration method for beam search on Transducer models, obtained by restructuring the beam itself as a batched data structure rather than a list of independent hypotheses. Each hypothesis is stored as a back-pointer chain in shared tensors, so extending a hypothesis costs one pointer update instead of copying a string, and two transcripts are compared for merging by an incremental polynomial hash in constant time. On top of this, the authors implement two fully batched decoders—ALSD++, a length-synchronous decoder, and AES++, a frame-synchronous decoder—and add CUDA-graph capture to eliminate kernel-launch overhead. They report that ALSD++ with an n-gram LM runs within 10–20% of greedy decoding for the whole system on three public datasets, gives 14–30% relative WER gains, and that their blank-scoring rule yields up to 11% relative improvement over existing shallow-fusion implementations.

Load-bearing premise

The speed and accuracy gains assume that the polynomial hash used to compare hypotheses almost never mistakes two different transcripts for the same one, a collision rate the paper puts at under 0.1% in its own tests but does not fully guard against.

Editorial extensions

If this is right

  • If the speed claim holds, streaming ASR systems can run beam search with an external language model at essentially the same cost as greedy decoding, which removes the main blocker for context biasing in production.
  • Low-resource ASR stands to gain 14–30% relative WER from language-model fusion at near-greedy speed, making beam search the default choice rather than a special mode.
  • Since the same batched trie and blank-scoring machinery works for TDT models, the acceleration is not tied to the RNN-T decoder and should survive as transducer variants evolve.
  • The ablation results imply that the practical value of the method grows with batch size: at larger batches the encoder dominates, so the remaining decoder overhead is hidden.

Reading between the lines

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

  • The paper does not test this, but the same incremental-hash and trie equality mechanism should transfer to CTC prefix beam search, where hypotheses also collapse modulo blank placement.
  • The paper's decoder-only measurements imply a caveat: the 10–20% whole-system gap is a batched-server result, and single-stream or encoder-light settings will see a larger beam-vs-greedy gap.
  • The paper does not explore this, but the blank-scoring rule generalizes to any sequence model with a 'no-output' symbol plus an external rescorer that lacks that symbol.
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

3 major / 5 minor

Summary. The paper proposes a GPU-oriented acceleration method for beam search decoding of Transducer-based ASR models. The method combines batched hypothesis processing, a tree-based transcript storage structure, hash-based constant-time hypothesis merging, CUDA graph execution, and a new blank-scoring scheme for shallow fusion with n-gram language models. It is instantiated in two algorithms, ALSD++ and AES++, and evaluated on SPGISpeech, Europarl, and SLURP with RNN-T and TDT models. The authors report beam search speeds within 10-20% of greedy decoding and 14-30% relative WER improvements with LM fusion, with code made available through a NeMo pull request.

Significance. If the reported results are reproducible, this is a practically valuable contribution: it substantially lowers the computational barrier to using beam search and LM shallow fusion in production Transducer ASR systems. The paper includes useful ablations over batch size, beam size, and four LM fusion strategies, and it builds on and clarifies earlier ALSD and AES approaches. The open-source implementation is a concrete strength, and the use of public datasets and a GPU-accelerated n-gram LM makes the experimental setup largely transparent.

major comments (3)
  1. [Section 2.1, Eq. (1)] The incremental polynomial hash is load-bearing for the claimed constant-time hypothesis merging, but the manuscript does not report the chosen base P and modulus M, the denominator of the claimed 'less than 0.1%' hash collision rate, or the residual false-merge rate after the additional length and latest-token checks. A false merge of two distinct transcripts with equal length and equal final token would silently replace a valid hypothesis and could degrade WER. Please report the parameter values, the collision statistics over the actual decoding runs, and the exact collision-handling logic, so that the reader can verify that ALSD++ and AES++ preserve exact beam search quality.
  2. [Section 2.3, Eq. (3)] The text states that the proposed blank scoring 'preserves total probability,' but the scores defined in Eq. (3) are not normalized: the sum over the vocabulary and the blank symbol is not equal to 1 in general and depends on lambda and on the correlation between p[k] and p_LM[k]. Please either specify precisely what quantity is preserved, or give the normalized scoring formula actually used in the implementation. This matters because Figure 2 and the fusion-strategy comparisons rely on Eq. (3) as a defined method.
  3. [Abstract and Section 4] The abstract's claim of '14-30% relative improvement in WER compared to greedy decoding' is not supported by Table 1 for the high-resource in-domain rows: for RNN-T on Europarl the improvement is 4.8%, and for TDT on Europarl it is 4.3%. Please qualify the claim to specify the setting (e.g., low-resource or out-of-domain test sets) or report the subset of results that supports the stated range.
minor comments (5)
  1. [Table 1] The column header 'no LM with LM' is ambiguous, and the ALSD row appears to have no LM-fusion entries while the table layout suggests it does. Please restructure the table with clearly separated 'no LM' and 'with LM' column groups.
  2. [Section 2.2] The parameter s ('maximum of s expansions per frame') is not defined in the text and no chosen value is reported in the experiments. Please define it and give the values used in the evaluations.
  3. [Section 3] RTFx values are reported as averages over three runs after a single warm-up, but no variance or standard deviation is given. Given the small differences in some comparisons (e.g., Table 2, batch size 32, RNN-T full-model greedy 228 vs beam 194), please report run-to-run variability or confidence intervals.
  4. [Section 3] The text says performance measurements use 'double-precision floating-point numbers.' Please clarify whether this means the model computations themselves are executed in double precision or only the timing/accumulation code, since this affects comparability with standard inference practices.
  5. [Abstract and Section 1] The open-source reference is a pull request (PR #12729) rather than a stable release. Please cite a released version, tag, or DOI so that the described algorithms and parameters can be retrieved unambiguously.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: the headline speed and WER figures are measured on public test sets, LM scores are fixed by an external ARPA file, and the self-citations (NGPU-LM, label-looping) are corroborated in-paper; the undisclosed hash-collision parameters in Sec. 2.1 are a verifiability gap, not a circular step.

full rationale

The paper's central claims — the 10–20% full-system speed gap, the 14–30% relative WER gain over greedy decoding, and the up-to-11% shallow-fusion gain — are reported measurements in Tables 1–3 on public test sets (SPGISpeech, SLURP, Europarl), not quantities derived from the method's own definitions. The only tuned parameter, the LM weight λ, is disclosed as tuned on development sets with test WER reported separately, so no fitted input is renamed as a prediction. Equation (3) is a proposed rescoring rule whose benefit (preventing deletion-rate growth at high λ) is verified by the measured WER-vs-λ curves in Fig. 2 rather than by construction; the paper's phrasing 'preserves total probability' is not literally satisfied for λ>0, which is a rigor quibble, not circularity. Equation (1) is a standard incremental polynomial hash, and the claim that collisions are 'less than 0.1% in practice' is an empirical observation with P and M undisclosed — a verifiability gap in a load-bearing assumption (Sec. 2.1), but not a reduction of the output to the input. The self-citations to NVIDIA prior work ([14] label-looping as the batching base, [22] NGPU-LM for batched LM queries) are present, but NGPU-LM's outputs are guaranteed identical to the external KenLM baseline by the shared ARPA file ('Both models have the same underlying ARPA format file, guaranteeing identical LM outputs'), so the late-pruning WER gains do not reduce to the self-citation, and the integrated RTFx with LM is measured in this paper (Table 1). The paper is therefore self-contained against external benchmarks; the missing hash parameters are a robustness concern, and the abstract's 14–30% range selectively excludes in-domain improvements of 4.3–4.8%, but neither issue constitutes circularity.

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

The headline claims depend on tunable lambda, beam size, expansion limit s, and hash constants, all of which the paper leaves partly unspecified. The contribution is empirical engineering, so these are free parameters or design assumptions rather than derived constants.

free parameters (4)
  • lambda, LM fusion weight = not reported; tuned per dataset on dev sets
    Controls the ASR/LM trade-off in Eqs. (2) and (3); all LM-fusion WER numbers depend on it and Figure 2 shows strong sensitivity.
  • beam size = 6 for main tables; 1-12 in Table 3
    Beam width drives both WER and speed; the headline 10-20% gap is reported at beam 6 and batch 32.
  • s, maximum expansions per frame = not specified
    ALSD++ and AES++ restrict expansions per frame, but the paper never reports s, so readers cannot tell how close the search is to the unconstrained beam search.
  • hash base P and modulus M = not reported
    The collision rate in Eq. (1) depends on these chosen constants and on token statistics; without them the '<0.1%' collision observation cannot be reproduced.
assumptions (4)
  • domain assumption Incremental polynomial rolling hash with a large prime modulus makes collisions negligible for transcripts produced by Transducer beam search.
    Invoked in Section 2.1 Eq. (1); underpins constant-time merging and the claimed linear decoding complexity.
  • domain assumption Replacing discarded beam hypotheses with additional expansions preserves the behavior of the original beam search.
    Section 2.1 states that exactly Beam hypotheses are maintained by substituting discarded hypotheses with extra expansions; this changes pruning semantics unless the extra expansions are the ones the original algorithm would keep.
  • domain assumption CUDA Graphs can capture the entire beam search iteration because tensor shapes and control flow are static.
    Section 2.1 relies on CUDA graph replay; if any data-dependent shape or branch exists, the graph must be re-captured and the speed claim would not hold.
  • domain assumption An n-gram LM trained on the training-set text is a fair external domain model, and NGPU-LM reproduces KenLM probabilities exactly.
    Section 3 compares against KenLM and claims identical ARPA outputs; the LM fusion gains depend on this equivalence and on the LM matching the test domain.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Pushing the Limits of Beam Search Decoding for Transducer-based ASR models." pith.science (2026). https://pith.science/paper/7I7JBVHI

@misc{pith2026250600185,
  author       = {Pith},
  title        = {Pith review of: Pushing the Limits of Beam Search Decoding for Transducer-based ASR models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/7I7JBVHI}},
  note         = {Machine review of arXiv:2506.00185}
}
read the original abstract

Transducer models have emerged as a promising choice for end-to-end ASR systems, offering a balanced trade-off between recognition accuracy, streaming capabilities, and inference speed in greedy decoding. However, beam search significantly slows down Transducers due to repeated evaluations of key network components, limiting practical applications. This paper introduces a universal method to accelerate beam search for Transducers, enabling the implementation of two optimized algorithms: ALSD++ and AES++. The proposed method utilizes batch operations, a tree-based hypothesis structure, novel blank scoring for enhanced shallow fusion, and CUDA graph execution for efficient GPU inference. This narrows the speed gap between beam and greedy modes to only 10-20% for the whole system, achieves 14-30% relative improvement in WER compared to greedy decoding, and improves shallow fusion for low-resource up to 11% compared to existing implementations. All the algorithms are open sourced.

Figures

Figures reproduced from arXiv: 2506.00185 by the authors.

Figure 1
Figure 1. Tree structure for batched hypotheses. non-blank token. Specifically, given an existing hash Ht for a hypothesis of length t, the hash for the extended hypothesis Ht+1 is computed as: Ht+1 = (Ht × P + Tt+1) (mod M) (1) where P is a chosen prime base, M is a large prime mod￾ulus, and Tt+1 is the newly appended token. For sufficiently large P and M, hash collisions are extremely rare (we ob￾serve less than 0.1% in pra… view at source ↗
Figure 2
Figure 2. WER vs λ (LM weight) for four LM fusion strategies. Model: RNN-T (∼ 1k). Data: SPGI dev. Strategy: ALSD++. Beam size: 6. to less than 20% for traditional Transducer models and under 10% for TDT models. High-resource: ASR models trained on large datasets can still benefit from external LM fusion when adapting to an ”un￾seen” domain. For RNN-T, it achieves a 17.8–19.4% improve￾ment on new domains and 4.8% in-domain. F… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

34 extracted references · 29 canonical work pages

  1. [1]

    Moreover, Transducer-based models are the pre- ferred choice for streaming ASR due to their causal decoders

    Introduction The Recurrent Neural Transducer (RNN-T or Transducer) ar- chitecture [1] is widely employed in Automatic Speech Recog- nition (ASR) systems, as it provides a balanced trade-off be- tween recognition accuracy and inference speed during greedy decoding. Moreover, Transducer-based models are the pre- ferred choice for streaming ASR due to their ...

    work page

  2. [2]

    next frame transition

    Methods A beam decoding iteration typically involves several key sub- steps. First, in the expansion step, current hypotheses within the beam are extended with one or more new tokens. Next, in the merging step, hypotheses with identical token sequences are combined. Finally, pruning is applied to retain the topN highest-scoring hypotheses for the next ste...

    work page 1932

  3. [3]

    In the low-resource scenario, we use a Fast Conformer Transducer Large model with around 115M parameters trained on 960 hours of LibriSpeech

    Experimental Setup ASR models:We evaluate low-resource and high-resource se- tups. In the low-resource scenario, we use a Fast Conformer Transducer Large model with around 115M parameters trained on 960 hours of LibriSpeech. In the high-resource scenario, we use RNN-T [23] and TDT [24] of similar architecture and size, trained on 24,000 hours of diverse E...

    work page 1919

  4. [4]

    Results A summary of our results is shown in Tab. 1. Both AES and ALSD offer minor improvements in WER, when decoded with- out external LM. AES++ and ALSD++ achieve similar accu- racy to their baselines while delivering significantly faster per- formance. ALSD++ outperforms AES++ in terms of speed, due less frequent prediction network updates. Low-resourc...

    work page

  5. [5]

    In this paper, we introduced a univer- sal beam search acceleration method that enables two novel fast decoding algorithms, ALSD++ and AES++

    Conclusion Beam search decoding has traditionally been a bottleneck for Transducer models, significantly slowing down inference com- pared to greedy decoding. In this paper, we introduced a univer- sal beam search acceleration method that enables two novel fast decoding algorithms, ALSD++ and AES++. Our approach in- tegrates batched processing, a tree-bas...

    work page

  6. [6]

    Alignment-Length Syn- chronous Decoding for RNN Transducer,

    G. Saon, Z. T ¨uske, and K. Audhkhasi, “Alignment-Length Syn- chronous Decoding for RNN Transducer,”ICASSP, 2020

    work page 2020

  7. [7]

    Sequence transduction with recurrent neural net- works,

    A. Graves, “Sequence transduction with recurrent neural net- works,” inICML, 2012

    2012

  8. [8]

    Shallow-fusion end-to-end contextual biasing,

    D. Zhao, T. N. Sainath, D. Rybach, P. Rondon, D. Bhatia, B. Li, and R. Pang, “Shallow-fusion end-to-end contextual biasing,” in Interspeech, 2019

    2019

Show all 34 references
  1. [9]

    Fast context-biasing for CTC and Transducer ASR models with CTC-based word spotter,

    A. Andrusenko, A. Laptev, V . Bataev, V . Lavrukhin, and B. Gins- burg, “Fast context-biasing for CTC and Transducer ASR models with CTC-based word spotter,”Interspeech, 2024

    2024

  2. [10]

    Acoustic look-ahead for more efficient decoding in LVCSR,

    D. Nolden, R. Schl ¨uter, and H. Ney, “Acoustic look-ahead for more efficient decoding in LVCSR,” inInterspeech, 2011

    2011

  3. [11]

    RNN-T for latency controlled ASR with improved beam search,

    M. Jain, K. Schubert, J. Mahadeokaret al., “RNN-T for latency controlled ASR with improved beam search,”ArXiv/1911.01629, 2019

    1911 arXiv

  4. [12]

    Efficient sequence transduction by jointly predicting tokens and durations,

    H. Xu, F. Jia, S. Majumdar, H. Huang, S. Watanabe, and B. Gins- burg, “Efficient sequence transduction by jointly predicting tokens and durations,”ICML, 2023

    2023

  5. [13]

    De- spite their promise to reduce beam search time, their speeds still remain significantly lower than greedy decoding

    to accelerate beam search decoding in streaming ASR. De- spite their promise to reduce beam search time, their speeds still remain significantly lower than greedy decoding. Fast greedy search methods for Transducers have further di- minished the need for beam search decoding. ...

    2025 arXiv

  6. [14]

    Accelerating RNN Transducer in- ference via adaptive expansion search,

    J. Kim, Y . Lee, and E. Kim, “Accelerating RNN Transducer in- ference via adaptive expansion search,”IEEE Signal Processing Letters, 2020

    2020

  7. [15]

    A token-wise beam search algorithm for RNN-T,

    G. Keren, “A token-wise beam search algorithm for RNN-T,” ASRU, 2023

    2023

  8. [16]

    Fast and parallel decoding for trans- ducer,

    W. Kang, L. Guo, F. Kuang, L. Lin, M. Luo, Z. Yao, X. Yang, P. ˙Zelasko, and D. Povey, “Fast and parallel decoding for trans- ducer,” inICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2023, pp. 1–5

    2023

  9. [17]

    RNN-Transducer with stateless prediction network,

    M. R. Ghodsi, X. Liu, J. A. Apfel, R. Cabrera, and E. Weinstein, “RNN-Transducer with stateless prediction network,”ICASSP, 2020

    2020

  10. [18]

    Accelerating RNN-T training and inference using CTC guidance,

    Y . Wang, Z. Chen, C. yong Zheng, Y . Zhang, W. Han, and P. Haghani, “Accelerating RNN-T training and inference using CTC guidance,”ICASSP, 2022

    2022

  11. [19]

    Streaming parallel transducer beam search with fast-slow cascaded encoders,

    J. Mahadeokar, Y . Shi, K. Li, D. Le, J. Zhu, V . Chandra, O. Kalinli, and M. L. Seltzer, “Streaming parallel transducer beam search with fast-slow cascaded encoders,” inInterspeech, 2022

    2022

  12. [20]

    Label-looping: Highly efficient decoding for transducers,

    V . Bataev, H. Xu, D. Galvez, V . Lavrukhin, and B. Ginsburg, “Label-looping: Highly efficient decoding for transducers,” in SLT, 2024

    2024

  13. [21]

    Speed of light exact greedy decoding for RNN-T speech recognition models on GPU,

    D. Galvez, V . Bataev, H. Xu, and T. Kaldewey, “Speed of light exact greedy decoding for RNN-T speech recognition models on GPU,”Interspeech, 2024

    2024

  14. [22]

    Nemo: a toolkit for build- ing ai applications using neural modules,

    O. Kuchaiev, J. Li, H. Nguyen, O. Hrinchuk, R. Leary, B. Gins- burg, S. Kriman, S. Beliaev, V . Lavrukhin, J. Cook, P. Castonguay, M. Popova, J. Huang, and J. M. Cohen, “Nemo: a toolkit for build- ing ai applications using neural modules,” 2019

    2019

  15. [23]

    ESPnet: End-to-end speech processing toolkit,

    S. Watanabe, T. Hori, S. Karita, T. Hayashi, J. Nishitoba, Y . Unno, N. Enrique Yalta Soplin, J. Heymann, M. Wiesner, N. Chen, A. Renduchintala, and T. Ochiai, “ESPnet: End-to-end speech processing toolkit,” inInterspeech, 2018

    2018

  16. [24]

    Connectionist temporal classification: labelling unsegmented se- quence data with recurrent neural networks,

    A. Graves, S. Fern ´andez, F. J. Gomez, and J. Schmidhuber, “Connectionist temporal classification: labelling unsegmented se- quence data with recurrent neural networks,”ICML, 2006

    2006

  17. [25]

    Neural machine transla- tion of rare words with subword units,

    R. Sennrich, B. Haddow, and A. Birch, “Neural machine transla- tion of rare words with subword units,” inProceedings of the 54th Annual Meeting of the Association for Computational Linguistics, 2016

    2016

  18. [26]

    Hybrid au- toregressive transducer (hat),

    E. Variani, D. Rybach, C. Allauzen, and M. Riley, “Hybrid au- toregressive transducer (hat),”ICASSP, 2020

    2020

  19. [27]

    Modular hybrid autoregressive transducer,

    Z. Meng, T. Chen, R. Prabhavalkar, and other, “Modular hybrid autoregressive transducer,”SLT, 2022

    2022

  20. [28]

    NGPU-LM: GPU-Accelerated N-Gram Lan- guage Model for context-biasing in greedy ASR decoding,

    V . Bataev, A. Andrusenko, L. Grigoryan, A. Laptev, V . Lavrukhin, and B. Ginsburg, “NGPU-LM: GPU-Accelerated N-Gram Lan- guage Model for context-biasing in greedy ASR decoding,” ArXiv/2505.22857, 2025

    2025 arXiv

  21. [29]

    STT En FastConformer Transducer Large,

    NVIDIA, “STT En FastConformer Transducer Large,” 2020. [Online]. Available: https://hf.co/nvidia/stt en fastconformer transducer large

    2020

  22. [30]

    STT En FastConformer TDT Large,

    NVIDIA, “STT En FastConformer TDT Large,” 2020. [Online]. Available: https://hf.co/nvidia/stt en fastconformer tdt large

    2020

  23. [31]

    SPGISpeech: 5,000 hours of transcribed financial audio for fully formatted end- to-end speech recognition,

    P. K. O’Neill, V . Lavrukhin, S. Majumdar, V . Noroozi, Y . Zhang, O. Kuchaiev, J. Balam, Y . Dovzhenko, K. Freyberg, M. D. Shul- man, B. Ginsburg, S. Watanabe, and G. Kucsko, “SPGISpeech: 5,000 hours of transcribed financial audio for fully formatted end- to-end speech recogn...

    2021 arXiv

  24. [32]

    SLURP: A Spoken Language Understanding Resource Package,

    E. Bastianelli, A. Vanzo, P. Swietojanski, and V . Rieser, “SLURP: A Spoken Language Understanding Resource Package,” in EMNLP, 2020

    2020

  25. [33]

    Europarl- ASR: A Large Corpus of Parliamentary Debates for Streaming ASR Benchmarking and Speech Data Filtering/Verbatimization,

    G. V . Garc ´es D ´ıaz-Mun´ıo, J. A. Silvestre-Cerd `a, J. Jorge, A. Gim ´enez, J. Iranzo-S ´anchez, P. Baquero-Arnal, N. Rosell ´o, A. P.-G. de Martos, J. Civera, A. Sanchis, and A. Juan, “Europarl- ASR: A Large Corpus of Parliamentary Debates for Streaming ASR Benchmarking a...

    2021

  26. [34]

    KenLM: Faster and smaller language model queries,

    K. Heafield, “KenLM: Faster and smaller language model queries,” inProc. of the Sixth Workshop on Statistical Machine Translation, 2011

    2011

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.