Pith. sign in

REVIEW 1 major objections 4 minor 53 references

A modular, open speech-to-speech translation system can preserve the source speaker's voice as well as commercial APIs, at the cost of higher latency and somewhat lower translation scores.

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 · deepseek-v4-flash

2026-08-01 17:40 UTC pith:P6VQPMY2

load-bearing objection A transparent, useful open baseline for cascaded S2ST; the multi-speaker alignment pipeline needs validation before Table 5 is trusted. the 1 major comments →

arxiv 2607.17544 v1 pith:P6VQPMY2 submitted 2026-07-20 eess.AS

X-Translator: A Real-Time Multilingual Speaker-Aware Speech-to-Speech Translation System

Yuxiang Zhao , Yichi Zhang , Yanjie An , Yanqiao Zhu , Zhanxun Liu , Yushen Chen , Qixi Zheng , Haina Zhu
show 5 more authors
Yunchong Xiao Keqi Deng Shuai Fan Kai Yu Xie Chen
This is my paper
classification eess.AS
keywords speech-to-speech translationcascaded systemstreaming ASRsegment commitmentspeaker prompt managementvoice preservationmulti-speaker translationreal-time latency
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 argues that a fully open, modular cascade of streaming ASR, machine translation, and prompt-conditioned TTS can serve as a deployable real-time speech-to-speech translation platform, provided it is wrapped in explicit runtime policies for segment commitment and speaker routing. The central claim is that these runtime policies let a low-cost system match or beat proprietary black-box translation services on speaker preservation and intelligibility of translated speech, even though it remains slower and slightly weaker on translation quality. A sympathetic reader should care because the result suggests the practical gap between open research systems and commercial translation products is narrower than model-centric comparisons imply, and because the design makes every delay or routing error inspectable.

Core claim

X-Translator shows that assembling streaming ASR, MT, and prompt-conditioned TTS behind a session-level controller yields a system whose translated speech keeps the source speaker's voice: in short-form evaluation it posts the highest speaker similarity in both directions (0.87 for English-to-Chinese and 0.88 for Chinese-to-English) and the lowest target word-error rate in Chinese-to-English (3.04%), while its start latency (2,584 and 4,056 ms) and COMET scores (0.78 and 0.74) trail the best proprietary APIs (814 and 800 ms; 0.81 and 0.77). In long-form single-speaker tests its target-voice consistency and source-speaker similarity match a top commercial API, and in multi-speaker sessions it

What carries the argument

The two load-bearing mechanisms are the incremental segment-commitment layer and the online speaker prompt manager. The commitment layer holds streaming ASR partial hypotheses as display-only text, cuts stable prefixes into translation-ready units using punctuation, silence, or no-update timeouts and speaker-boundary flushes, and sends only those committed segments to MT and TTS, preventing speech inflation from unstable hypotheses. The speaker prompt manager tracks speaker activity with speaker embeddings, assigns each committed segment to the speaker with the largest temporal overlap, maintains per-speaker fixed and rolling six-second prompt buckets, and routes the chosen prompt into TTS.

Load-bearing premise

The multi-speaker voice-preservation results assume that the automatically matched source and translated segments really are the same utterances; if that alignment is wrong, the reported speaker-similarity and purity scores measure the matching procedure rather than actual voice preservation.

What would settle it

Take a sample of the matched source-target segment pairs from the multi-speaker evaluation, have human annotators flag whether each pair is genuinely the same content, and recompute segment-level speaker similarity and purity on confirmed pairs; a material drop in either number would indicate the headline speaker-preservation claims are artifacts of the semantic-plus-percentile alignment.

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

If this is right

  • An open cascade can serve as a reproducible reference point for real-time speech-to-speech translation, with module-level logs that trace errors and delays to ASR commitment, translation, synthesis, speaker routing, or playback.
  • Because ASR, MT, and TTS are replaceable backends, the same runtime policies can absorb future component improvements without redesign, so the latency gap to proprietary systems may close as faster streaming models appear.
  • Speaker preservation in translated speech is achievable without end-to-end training, through prompt-conditioned synthesis supported by overlap-based speaker routing and dual fixed/rolling prompts.
  • Multi-speaker translation quality can be evaluated beyond sentence-level metrics, using segment-level speaker similarity together with target and source purity to separate speaker collapse from fragmentation.
  • The system's behavior can be compared behaviorally with opaque commercial APIs while retaining an internal evidence trail those APIs cannot provide.

Where Pith is reading between the lines

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

  • The same commitment-and-routing architecture could be transferred to live captioning, simultaneous interpreting, or assistive listening devices, where the key question is also when to commit a partial hypothesis and whose voice to keep.
  • The evaluation pipeline's semantic-plus-temporal segment matching could be repurposed as an automatic alignment tool for building speech-translation training data, but only after its matching accuracy is validated by human checks or known-match controls.
  • A testable extension would be an explicit commitment-threshold sweep: reporting quality and speaker similarity as a function of silence timeout and punctuation policy would turn the latency-quality trade-off into a tunable design curve rather than a single operating point.
  • The fixed-plus-rolling prompt strategy suggests a general recipe for long-session zero-shot voice cloning: a stable early reference anchors identity while a recent prompt tracks vocal drift, a combination worth testing outside translation services.

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

1 major / 4 minor

Summary. The paper presents X-Translator, an open modular cascaded speech-to-speech translation (S2ST) system that combines streaming ASR, machine translation, and prompt-conditioned TTS under a session-level runtime controller. The controller commits stable ASR segments and routes each segment to a speaker-specific voice prompt for synthesis. The authors evaluate the system on short-form OpenSTBench benchmarks, long-form single-speaker data (TEDLIUM-3 and WenetSpeech), multi-speaker conversations (VoxConverse and Bilibili interviews), and 70 FLEURS translation directions. They report that X-Translator achieves the highest speaker similarity and low target-speech error rates in short-form tests, but higher latency and slightly lower translation quality than proprietary APIs. The paper concludes that X-Translator is a practical, reproducible, low-cost alternative for deployment-oriented S2ST and provides a platform for studying runtime policies.

Significance. If the results hold, this is a valuable contribution: an open, inspectable S2ST system with explicit runtime policies and a level of documentation that is rare in the field. The manuscript is unusually transparent, documenting backend versions, serving modes, alignment rules, success-rate filtering, and baseline version drift in the appendices, and the code and evaluation artifacts are released. The short-form and long-form evaluations are carefully designed, with clear distinction between S2ST and S2TT metrics. However, the central speaker-preservation claim in multi-speaker settings relies on an unvalidated alignment pipeline, which currently prevents full confidence in the headline multi-speaker results. With that validation added, the paper would be a strong systems contribution.

major comments (1)
  1. [§4.4.3, Table 5, Eqs. (1)–(4)] The multi-speaker evaluation is the load-bearing support for the paper's speaker-awareness claim. Spk. SIM (Eq. 1) and the purity metrics (Eqs. 3–4) are computed after a two-stage alignment: VibeVoice-ASR diarization of the source and target audio, followed by DashScope semantic similarity matching with a midpoint-percentile constraint of 0.10. No validation of this alignment is reported—no human check, no known-match control, and no sensitivity analysis over the 0.10 threshold. Because the same matching defines the segment pairs and the speaker correspondence matrix C, any misalignment (wrong-speaker matches, many-to-one matches, or diarization errors) directly contaminates the reported Spk. SIM (0.77) and Target Purity (0.85/0.84). The concern is heightened because source and target languages, durations, and speaker counts differ. I request that the authors either provide evidence that
minor comments (4)
  1. [§4.4.1, Table 2] The metric cATD is introduced without a definition. The text says it is 'computed from our runtime logs,' yet Table 2 also reports cATD for proprietary baselines, which cannot have such logs. Please define cATD precisely, state how it is measured for each system, and justify its comparability to the externally measured ATD; otherwise, remove it and rely on Start and ATD, which already support the latency conclusions.
  2. [Tables 2 and 3] Baseline values are averaged over only two runs, with no spread or run-level detail reported. Given variability in proprietary APIs (some success rates drop below 94%), the mean of two runs could be unstable. Please provide run-level values, standard errors, or a statement on observed variance.
  3. [§4.4.2, Table 4] The long-form GSS metric compares each target-window embedding with a single normalized average source-speaker embedding. This average may be sensitive to source variability and duration normalization. A brief explanation of how the average is computed, or a per-window scatter or confidence interval, would aid interpretation.
  4. [§4.4.2 and Appendix D] Several thresholds are set without sensitivity analysis: γ=0.5 for long-gap similarity, the 10-s/5-s window/hop, and the multilingual coverage policy. While these are reasonable defaults, a short sensitivity check (e.g., γ∈{0.3,0.7}) would confirm that the qualitative conclusions are not threshold artifacts.

Circularity Check

0 steps flagged

No significant circularity: the paper reports measurements against external and modular baselines; the only self-referential elements (OpenSTBench, X-Voice) are not load-bearing.

full rationale

X-Translator is a systems paper: its claims are supported by direct measurements, not by a derivation chain that reduces to its inputs. The central metrics — Spk. SIM, Target/Source Purity, COMET, WER, CER, and latency — are defined explicitly (Eqs. 1–4) and computed from source/target audio or transcripts; none is a fitted parameter renamed as a prediction. The multi-speaker alignment pipeline (semantic matching plus a 0.10 midpoint-percentile constraint) is a validity risk because no validation of matching accuracy is reported, but it is not circular: the metrics are not defined in terms of the system's own routing decisions, and a misalignment would produce measurement error rather than a tautology. The paper does cite OpenSTBench (An et al., 2026), which is co-authored by the same group, and uses the authors' own X-Voice TTS, but the evaluation is self-contained against independent proprietary APIs and public datasets (TEDLIUM-3, WenetSpeech, VoxConverse, FLEURS), the metric formulas are stated in the paper, and the strongest claim explicitly concedes that latency and translation quality do not consistently surpass proprietary systems. There is no imported uniqueness theorem, no ansatz smuggled via self-citation, and no renaming of a known result as a prediction. The low score reflects minor self-referential artifacts rather than any load-bearing circular step.

Axiom & Free-Parameter Ledger

4 free parameters · 6 axioms · 1 invented entities

The paper introduces no new physical entities or fitted model parameters; its free parameters are hand-chosen evaluation thresholds that shape the reported numbers. The main assumptions concern whether proxy measurements (Resemblyzer similarity, Whisper transcription, semantic matching) faithfully represent voice preservation and translation quality, and whether the custom latency metric is comparable across systems.

free parameters (4)
  • Long-gap threshold γ = 0.5 (normalized progress gap)
    Hand-chosen cutoff for SSLG/TTLG in Section 4.4.2; no sensitivity analysis, and it directly decides which window pairs count as 'long-gap' for the long-form stability claim.
  • Multi-speaker matching constraint ε = 0.10 midpoint-percentile
    Hand-chosen search bound for semantic matching in Section 4.4.3; it controls the alignment that all multi-speaker Spk. SIM and purity numbers (Table 5) depend on, with no sensitivity check reported.
  • Prompt length = 6 seconds (fixed and rolling)
    Fixed by design in Section 3.3; prompt duration affects TTS voice quality and hence every speaker-similarity number; not swept.
  • Embedding window/hop for long-form metrics = 10 s window, 5 s hop; CAM++ 1.0 s window, 0.5 s hop
    Chosen in Section 4.4.2 and Appendix C; windowing affects TTLG/GSS and speaker-attribution results.
axioms (6)
  • domain assumption Cosine similarity between L2-normalized Resemblyzer embeddings measures speaker voice similarity for source and synthesized speech
    Used as Spk. SIM throughout (Sections 4.4.1–4.4.3). Resemblyzer is trained mostly on English telephone speech; transfer to Chinese, to other languages, and to cloned/synthesized voices is assumed, not validated here.
  • domain assumption Feeding complete long recordings continuously into the system simulates real-time streaming behavior
    Section 4.1: 'simulate streaming input by feeding each recording to the system continuously'; real network conditions, user listening, and interactive turn-taking are absent.
  • domain assumption FLEURS shared sentence IDs across languages are semantically aligned
    Section 4.1 and Appendix D: 'we evaluate only on samples with shared sentence IDs in both the source and target languages, ensuring semantically aligned and consistent test sets'.
  • domain assumption Whisper-medium transcription with an explicit target-language code is a faithful proxy for evaluating generated target speech in all 19 output languages
    Appendix D: multilingual COMET is computed on Whisper-medium transcripts; ASR errors enter the score and likely vary by language (std up to 0.09 in Table 6).
  • ad hoc to paper The custom cATD latency metric computed from X-Translator's internal logs is comparable to the externally observed ATD of proprietary baselines
    Section 4.4.1: 'cATD denotes the custom ATD variant computed from our runtime logs'; the paper does not specify how the same quantity is derived for black-box baselines, so the reported latency comparisons may not be apples-to-apples.
  • domain assumption The overlap rule binding committed ASR segments to the speaker with maximal temporal overlap yields correct speaker attribution
    Section 3.3: 'assigns the segment to the speaker ID with the largest overlap'; this heuristic is load-bearing for all speaker-aware routing but is not evaluated for accuracy.
invented entities (1)
  • Session-level runtime controller with incremental segment commitment and online speaker prompt manager independent evidence
    purpose: Converts unstable ASR streams into committed translation units and routes each unit to a speaker-specific TTS prompt
    A software construct, not a new physical entity. It is shipped in the released code and exposes runtime logs, so it is falsifiable by reproduction; however, its claimed benefit (stable long-form speaker-aware translation) is only evidenced by the paper's own measurements.

pith-pipeline@v1.3.0-alltime-deepseek · 17808 in / 22086 out tokens · 183812 ms · 2026-08-01T17:40:59.501716+00:00 · methodology

0 comments
read the original abstract

Real-time speech-to-speech translation (S2ST) systems must balance translation quality, latency, speech naturalness, and speaker consistency. Publicly documented S2ST systems have advanced direct, multilingual, streaming, and expressive modeling, while proprietary products and APIs increasingly expose real-time translation capabilities to users. However, practical deployment remains challenging for open and reproducible systems, especially in long-form and multi-speaker conversations where partial ASR hypotheses are unstable, turn boundaries are ambiguous, and target speech must be generated with an appropriate speaker prompt. We present X-Translator, a low-cost modular cascaded S2ST system that combines streaming ASR, machine translation, and prompt-conditioned TTS through a session-level runtime controller. The system uses incremental segment commitment to convert unstable ASR streams into translation-ready units, and an online speaker prompt manager to bind source speech spans to speaker-specific voice prompts for synthesis. We evaluate translation, speech quality, and latency with OpenSTBench, compare against proprietary speech translation APIs as behavioral baselines, measure long-form voice stability, evaluate speaker preservation in multi-speaker conversations, and assess multilingual translation quality. X-Translator provides an open platform for understanding the practical trade-offs of deployment-oriented S2ST. Code and demo are available at https://github.com/zhaoyx239/X-Translator.

Figures

Figures reproduced from arXiv: 2607.17544 by Haina Zhu, Kai Yu, Keqi Deng, Qixi Zheng, Shuai Fan, Xie Chen, Yanjie An, Yanqiao Zhu, Yichi Zhang, Yunchong Xiao, Yushen Chen, Yuxiang Zhao, Zhanxun Liu.

Figure 1
Figure 1. Figure 1: System architecture of X-Translator. The browser streams audio and session controls to the server through [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: ASR pipeline and segment commitment policy. The left panel shows that streaming ASR hypotheses are [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Speaker prompt manager. The left panel shows online speaker tagging over the streaming audio timeline. [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗

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

53 extracted references · 3 canonical work pages

  1. [1]

    The Computational Geometry Algorithms Library , author =

  2. [2]

    Menelaos Karavelas , subtitle =

  3. [3]

    The Computational Geometry Algorithms Library , subtitle =

    Menelaos Karavelas , editor =. The Computational Geometry Algorithms Library , subtitle =

  4. [4]

    The Parmap library , author =

  5. [5]

    Christopher Anderson and Sophia Drossopoulou , title =

  6. [6]

    arXiv preprint arXiv:1904.06037 , year =

    Direct Speech-to-Speech Translation with a Sequence-to-Sequence Model , author =. arXiv preprint arXiv:1904.06037 , year =

    Pith/arXiv arXiv 1904

  7. [7]

    arXiv preprint arXiv:2107.08661 , year =

    Translatotron 2: High-quality direct speech-to-speech translation with voice preservation , author =. arXiv preprint arXiv:2107.08661 , year =

    Pith/arXiv arXiv

  8. [8]

    arXiv preprint arXiv:2308.11596 , year =

    SeamlessM4T: Massively Multilingual and Multimodal Machine Translation , author =. arXiv preprint arXiv:2308.11596 , year =

    Pith/arXiv arXiv

  9. [9]

    arXiv preprint arXiv:2201.03713 , year =

    CVSS Corpus and Massively Multilingual Speech-to-Speech Translation , author =. arXiv preprint arXiv:2201.03713 , year =

    Pith/arXiv arXiv

  10. [10]

    2019 , address =

    Ma, Mingbo and Huang, Liang and Xiong, Hao and Zheng, Renjie and Liu, Kaibo and Zheng, Baigong and Zhang, Chuanqiang and He, Zhongjun and Liu, Hairong and Li, Xing and Wu, Hua and Wang, Haifeng , booktitle =. 2019 , address =. doi:10.18653/v1/P19-1289 , pages =

    doi:10.18653/v1/p19-1289 2019

  11. [11]

    arXiv preprint arXiv:1909.12406 , year =

    Monotonic Multihead Attention , author =. arXiv preprint arXiv:1909.12406 , year =

    Pith/arXiv arXiv 1909

  12. [12]

    arXiv preprint arXiv:2007.16193 , year =

    SimulEval: An Evaluation Toolkit for Simultaneous Translation , author =. arXiv preprint arXiv:2007.16193 , year =

    Pith/arXiv arXiv 2007

  13. [13]

    arXiv preprint arXiv:2110.08250 , year =

    Direct Simultaneous Speech-to-Speech Translation with Variational Monotonic Multihead Attention , author =. arXiv preprint arXiv:2110.08250 , year =

    Pith/arXiv arXiv

  14. [14]

    arXiv preprint arXiv:2406.03049 , year =

    StreamSpeech: Simultaneous Speech-to-Speech Translation with Multi-task Learning , author =. arXiv preprint arXiv:2406.03049 , year =

    Pith/arXiv arXiv

  15. [15]

    arXiv preprint arXiv:2301.02111 , year =

    Neural Codec Language Models are Zero-Shot Text to Speech Synthesizers , author =. arXiv preprint arXiv:2301.02111 , year =

    Pith/arXiv arXiv

  16. [16]

    arXiv preprint arXiv:2306.15687 , year =

    Voicebox: Text-Guided Multilingual Universal Speech Generation at Scale , author =. arXiv preprint arXiv:2306.15687 , year =

    Pith/arXiv arXiv

  17. [17]

    arXiv preprint arXiv:2407.05407 , year =

    CosyVoice: A Scalable Multilingual Zero-shot Text-to-speech Synthesizer based on Supervised Semantic Tokens , author =. arXiv preprint arXiv:2407.05407 , year =

    Pith/arXiv arXiv

  18. [18]

    arXiv preprint arXiv:2002.01320 , year =

    CoVoST: A Diverse Multilingual Speech-To-Text Translation Corpus , author =. arXiv preprint arXiv:2002.01320 , year =

    Pith/arXiv arXiv 2002

  19. [19]

    arXiv preprint arXiv:2205.12446 , year =

    FLEURS: Few-shot Learning Evaluation of Universal Representations of Speech , author =. arXiv preprint arXiv:2205.12446 , year =

    Pith/arXiv arXiv

  20. [20]

    arXiv preprint arXiv:2211.04508 , year =

    SpeechMatrix: A Large-Scale Mined Corpus of Multilingual Speech-to-Speech Translations , author =. arXiv preprint arXiv:2211.04508 , year =

    Pith/arXiv arXiv

  21. [21]

    arXiv preprint arXiv:2605.30792 , year =

    OpenSTBench: Beyond Semantic Evaluation for Speech Translation , author =. arXiv preprint arXiv:2605.30792 , year =

    Pith/arXiv arXiv

  22. [22]

    arXiv preprint arXiv:2107.05604 , year =

    Direct speech-to-speech translation with discrete units , author =. arXiv preprint arXiv:2107.05604 , year =

    Pith/arXiv arXiv

  23. [23]

    arXiv preprint arXiv:2112.08352 , year =

    Textless Speech-to-Speech Translation on Real Data , author =. arXiv preprint arXiv:2112.08352 , year =

    Pith/arXiv arXiv

  24. [24]

    arXiv preprint arXiv:2205.12523 , year =

    TranSpeech: Speech-to-Speech Translation With Bilateral Perturbation , author =. arXiv preprint arXiv:2205.12523 , year =

    Pith/arXiv arXiv

  25. [25]

    arXiv preprint arXiv:2310.07403 , year =

    DASpeech: Directed Acyclic Transformer for Fast and High-quality Speech-to-Speech Translation , author =. arXiv preprint arXiv:2310.07403 , year =

    Pith/arXiv arXiv

  26. [26]

    arXiv preprint arXiv:2312.05187 , year =

    Seamless: Multilingual Expressive and Streaming Speech Translation , author =. arXiv preprint arXiv:2312.05187 , year =

    Pith/arXiv arXiv

  27. [27]

    arXiv preprint arXiv:2507.17527 , year =

    Seed LiveInterpret 2.0: End-to-end Simultaneous Speech-to-speech Translation with Your Voice , author =. arXiv preprint arXiv:2507.17527 , year =

    Pith/arXiv arXiv

  28. [28]

    arXiv preprint arXiv:2509.21144 , year =

    UniSS: Unified Expressive Speech-to-Speech Translation with Your Voice , author =. arXiv preprint arXiv:2509.21144 , year =

    arXiv

  29. [29]

    Transactions of the Association for Computational Linguistics , volume =

    Multilingual Denoising Pre-training for Neural Machine Translation , author =. Transactions of the Association for Computational Linguistics , volume =. 2020 , doi =

    2020

  30. [30]

    arXiv preprint arXiv:2207.04672 , year =

    No Language Left Behind: Scaling Human-Centered Machine Translation , author =. arXiv preprint arXiv:2207.04672 , year =

    Pith/arXiv arXiv

  31. [31]

    Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing , month = nov, year =

    Cross Attention Augmented Transducer Networks for Simultaneous Translation , author =. Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing , month = nov, year =. doi:10.18653/v1/2021.emnlp-main.4 , pages =

    work page doi:10.18653/v1/2021.emnlp-main.4 2021

  32. [32]

    Interspeech 2020 , publisher =

    Low-Latency Sequence-to-Sequence Speech Recognition and Translation by Partial Hypothesis Selection , author =. Interspeech 2020 , publisher =. 2020 , pages =

    2020

  33. [33]

    Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , month = jul, year =

    Attention as a Guide for Simultaneous Speech Translation , author =. Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , month = jul, year =. doi:10.18653/v1/2023.acl-long.745 , pages =

    doi:10.18653/v1/2023.acl-long.745 2023

  34. [34]

    Interspeech 2023 , year =

    AlignAtt: Using Attention-based Audio-Translation Alignments as a Guide for Simultaneous Speech Translation , author =. Interspeech 2023 , year =

    2023

  35. [35]

    Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing , month = nov, year =

    Simul-MuST-C: Simultaneous Multilingual Speech Translation Corpus Using Large Language Model , author =. Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing , month = nov, year =. doi:10.18653/v1/2024.emnlp-main.1238 , pages =

    work page doi:10.18653/v1/2024.emnlp-main.1238 2024

  36. [36]

    Discourse Processes , volume =

    Turn-end Estimation in Conversational Turn-taking: The Roles of Context and Prosody , author =. Discourse Processes , volume =. 2021 , doi =

    2021

  37. [37]

    Interspeech 2021 , publisher =

    CoVoST 2 and Massively Multilingual Speech Translation , author =. Interspeech 2021 , publisher =. 2021 , pages =

    2021

  38. [38]

    F5- TTS : A Fairytaler that Fakes Fluent and Faithful Speech with Flow Matching

    Chen, Yushen and Niu, Zhikang and Ma, Ziyang and Deng, Keqi and Wang, Chunhui and Zhao, Jian and Yu, Kai and Chen, Xie. F5- TTS : A Fairytaler that Fakes Fluent and Faithful Speech with Flow Matching. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025

    2025

  39. [39]

    IEEE Access , volume=

    Enhancing automatic speech recognition with personalized models: Improving accuracy through individualized fine-tuning , author=. IEEE Access , volume=. 2024 , publisher=

    2024

  40. [40]

    ICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) , pages=

    Wenetspeech: A 10000+ hours multi-domain mandarin corpus for speech recognition , author=. ICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) , pages=. 2022 , organization=

    2022

  41. [41]

    Spot the Conversation: Speaker Diarisation in the Wild , url=

    Chung, Joon Son and Huh, Jaesung and Nagrani, Arsha and Afouras, Triantafyllos and Zisserman, Andrew , year=. Spot the Conversation: Speaker Diarisation in the Wild , url=. doi:10.21437/interspeech.2020-2337 , booktitle=

    work page doi:10.21437/interspeech.2020-2337 2020

  42. [42]

    2026 , eprint=

    VIBEVOICE-ASR Technical Report , author=. 2026 , eprint=

    2026

  43. [43]

    2025 , eprint=

    Qwen3 Embedding: Advancing Text Embedding and Reranking Through Foundation Models , author=. 2025 , eprint=

    2025

  44. [44]

    arXiv preprint arXiv:2303.00332 , year=

    Cam++: A fast and efficient network for speaker verification using context-aware masking , author=. arXiv preprint arXiv:2303.00332 , year=

    Pith/arXiv arXiv

  45. [45]

    arXiv preprint arXiv:2601.21337 , year=

    Qwen3-ASR Technical Report , author=. arXiv preprint arXiv:2601.21337 , year=

    Pith/arXiv arXiv

  46. [46]

    arXiv preprint arXiv:2206.08317 , year=

    Paraformer: Fast and Accurate Parallel Transformer for Non-autoregressive End-to-End Speech Recognition , author=. arXiv preprint arXiv:2206.08317 , year=

    Pith/arXiv arXiv

  47. [47]

    arXiv preprint arXiv:2407.04051 , year=

    FunAudioLLM: Voice Understanding and Generation Foundation Models for Natural Interaction Between Humans and LLMs , author=. arXiv preprint arXiv:2407.04051 , year=

    Pith/arXiv arXiv

  48. [48]

    2025 , eprint=

    NiuTrans.LMT: Toward Inclusive and Scalable Multilingual Machine Translation with LLMs , author=. 2025 , eprint=

    2025

  49. [49]

    arXiv preprint arXiv:2512.24092 , year=

    HY-MT1.5 Technical Report , author=. arXiv preprint arXiv:2512.24092 , year=

    arXiv

  50. [50]

    arXiv preprint arXiv:2502.05512 , year=

    IndexTTS: An Industrial-Level Controllable and Efficient Zero-Shot Text-To-Speech System , author=. arXiv preprint arXiv:2502.05512 , year=

    Pith/arXiv arXiv

  51. [51]

    arXiv preprint arXiv:2605.05611 , year=

    X-Voice: Enabling Everyone to Speak 30 Languages via Zero-Shot Cross-Lingual Voice Cloning , author=. arXiv preprint arXiv:2605.05611 , year=

    Pith/arXiv arXiv

  52. [52]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume=

    Indextts2: A breakthrough in emotionally expressive and duration-controlled auto-regressive zero-shot text-to-speech , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume=

  53. [53]

    International conference on speech and computer , pages=

    TED-LIUM 3: Twice as much data and corpus repartition for experiments on speaker adaptation , author=. International conference on speech and computer , pages=. 2018 , organization=

    2018