Typhoon ASR Streaming: Steerable Low-Latency Thai Speech Recognition with Real-Time Shallow Fusion
Warit Sirichotedumrong, Tanawin Samutsin, Shah Faisal Wani, Sittipong Sripaisarnmongkol, Kunat Pipatanakul
cs.CL · cs.SD · eess.AS
Quick verdict
Deployable streaming Thai ASR: cache-aware models cut streamed CER 4.3-4.5x and n-gram fusion lifts keyword recall ~4pts; solid but single-set, modest gains.
Summary
The paper presents a deployable streaming Thai ASR system built on FastConformer-Transducer architectures. The authors show that their previously released full-context Typhoon ASR model collapses when forced to decode a true audio stream (CER rising from 11.0% offline to 62.8% at 1040 ms look-ahead and 100% at 80 ms). They restore streaming viability through two routes: converting the full-context 115M checkpoint to a cache-aware encoder plus brief fine-tuning, and adapting NVIDIA's Nemotron 0.6B streaming model to Thai by extending its tokenizer with Thai BPE plus numeric fine-tuning. The 0.6B model reaches 14.1% CER on TVSpeech and 9.3% on GigaSpeech2-Thai at 1040 ms look-ahead — a 4.3–4.5× reduction versus the streamed full-context model — while running faster than real time (RTF 0.024, first token ~1.06 s). A shallow-fusion layer re-ranks transducer candidates with a GPU 4-gram LM and a phrase-boosting tree, lifting English code-switch keyword recall on a 285-utterance held-out TVSpeech half from 16.6% to 20.4–20.7% at unchanged CER and under 3% RTF overhead. Ablations attribute most of the gain to a general n-gram over ordinary training transcripts (20.2% alone), which fixes inconsistent spellings of code-switched words; an upper-bound n-gram trained on test references reaches 27.8%. Limitations: steering evidence is confined to one benchmark, one model, and one latency, with no statistical significance testing, and steering resources are partly derived from the evaluation set's references.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 3
● low 2
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 18 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- A full-context Thai FastConformer-Transducer collapses under true streaming (CER 11.0% → 62.8% at 1040 ms look-ahead, 100% at 80 ms), and cache-aware conversion or adaptation restores it to 14.1%/9.3% CER on TVSpeech/GigaSpeech2-Thai at 1040 ms — a 4.3–4.5× reduction at faster-than-real-time speed.
- Decode-time shallow fusion with a GPU 4-gram LM and phrase boosting lifts English code-switch keyword recall on TVSpeech's held-out half from 16.6% to 20.4–20.7% at essentially unchanged CER and under 3% RTF overhead.
- Most of the recall gain comes from a general n-gram trained only on ordinary ASR transcripts (20.2% alone), which pins inconsistently spelled code-switched words to canonical written forms; keyword synthesis and phrase boosting matter mainly for rare terms.
- The steering ceiling is 27.8% recall (an n-gram trained on test references, non-deployable), indicating substantial headroom for a stronger in-domain LM.
Strengths
Quantified collapse diagnosis
Section IV-B / Fig. 2 measures exactly how the full-context model degrades under streaming (11.0% → 62.8% CER at 1040 ms, 100% at 80 ms) across two benchmarks and four look-ahead settings, turning an anecdote into a reproducible motivation number.
Honest ceiling row
Table II includes a non-deployable upper bound — an n-gram trained on the test references (27.8% recall, 10.2% CER) — explicitly framed as a memorization ceiling; transparent hygiene rarely seen in fusion papers.
Non-circular ablation isolates the insight
Deployment-grade efficiency reporting
Full artifact release
Weaknesses
Modest steering gain on a small, single-domain test
No statistical uncertainty anywhere
Steering resources partly derived from the evaluation set
Thin baseline coverage
'Model-agnostic' overclaim
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
This IEEE-conference-style system paper from SCB 10X's Typhoon team attacks two coupled gaps in open Thai ASR: (1) the ecosystem is dominated by offline Whisper-style models unusable for live captioning or voice agents, and (2) vocabulary control (names, domain terms, code-switched English) conventionally requires per-customer retraining. Three claims are advanced: a full-context FastConformer-Transducer collapses under true streaming (CER 11.0% offline → 62.8% at 1040 ms look-ahead, 100% at 80 ms; Section IV-B, Fig. 2); cache-aware conversion or adaptation restores streaming (0.6B model: 14.1% CER on TVSpeech and 9.3% on GigaSpeech2-Thai at 1040 ms, a 4.3–4.5× cut, faster than real time); and shallow fusion (Eq. 1: GPU 4-gram LM plus phrase-boost tree) steers vocabulary at decode time, lifting English code-switch keyword recall from 16.6% to 20.4–20.7% at unchanged CER and under 3% RTF overhead. The authors are explicit that they package a known technique for Thai rather than invent one (Section I), which sets the novelty bar at systemization plus Thai-specific evidence — a bar the paper mostly meets.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What accuracy does Typhoon ASR Streaming achieve in real time?
With the 0.6B streaming model at a one-second (1040 ms) look-ahead, the paper reports 14.1% character error rate (CER) on TVSpeech (code-switched Thai news/politics speech) and 9.3% on GigaSpeech2-Thai (general YouTube speech), versus 62.8% and 39.8% for the previous full-context model forced to stream. It runs faster than real time: batch-1 RTF 0.024 on one NVIDIA H100, with the first token emitted after about 1.06 seconds.
How does the keyword boosting work without retraining the ASR model?
At each decoding step, the system re-scores the transducer's candidate next tokens by adding two terms to the acoustic score: a weighted GPU n-gram language model score and a bonus for tokens along a phrase-boosting tree built from a user-supplied word list. Model weights never change. The paper uses equal weights (0.5 each) and measures under 3% real-time-factor overhead on an H100.
Are the models and code publicly available?
Yes, per the paper: the 115M and 0.6B streaming checkpoints, the fusion and streaming code, and the Thai evaluation suite are released through the project page (warit-s.github.io/typhoon-asr-streaming), built on NVIDIA NeMo. The evaluation sets (TVSpeech, GigaSpeech2-Thai) come from the public Typhoon ASR Benchmark, so the reported CER and keyword-recall numbers can in principle be re-run.
Why does the paper report character error rate instead of word error rate for Thai?
Thai is written without spaces between words, so word boundaries are ambiguous and word error rate becomes ill-defined. The paper first normalizes transcripts to one canonical written form (resolving numbers, the mai yamok repetition marker, and loanword spellings) and then reports character error rate (CER), the standard metric for space-free scripts. It additionally tracks keyword recall for English code-switched terms on the TVSpeech set.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Fig. 2's per-series values were extracted as flat number runs; mapping value series to legend entries is ambiguous (the naive order conflicts with the text's claim that the 0.6B leads at every latency), so per-latency values for the 115M vs 0.6B — especially at 80 ms — are uncertain.
- Figure 1 (system diagram) is assessable only via its caption; no visual verification possible.
- IEEE conference-format paper: no appendix or extended ablations exist beyond the extracted text; fusion is evaluated only in the reported single configuration.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.