Design of the IBM Granite 5.0 TurboCTC ASR Model
Brian Kingsbury, George Saon, Masayuki Suzuki, Hong-Kwang J. Kuo, Takashi Fukuda, Samuel Thomas, Vishal Sunder, Avihu Dekel
cs.CL
Quick verdict
Claims a 470M CTC Conformer on the OpenASR Pareto frontier at 2x rival speed; ablations broadly support it, but ~0.1% WER gains lack variance or significance analysis.
Summary
This paper presents Granite 5.0 Turbo CTC, a 470-million-parameter encoder-only CTC Conformer for fast, accurate English ASR. The architecture combines pyramidal temporal subsampling implemented as strided depthwise convolutions in the first two Conformer blocks, block-diagonal (chunk-wise) self-attention with 128-frame chunks, and conditioning on intermediate predictions from the eighth layer. Training uses roughly 60k hours of public and synthetic data with balanced sampling, a Muon optimizer with Polar Express orthogonalization (outperforming AdamW: 5.83% vs 5.99% average WER), knowledge distillation from the granite-4.1-8b-base LLM through a discarded 75M auxiliary decoder, and robust fine-tuning on clean/noisy paired views with parameter interpolation. Replacing 1x1 convolutions with linear layers and switching to an efficient attention kernel yields a 49% inverse-RTFx speedup. The model attains 5.83% average WER on the public Open ASR leaderboard test sets, sits on the leaderboard's speed-accuracy Pareto frontier as of 4 September 2026, and runs about twice as fast as the nearest competitor; the companion -nc model was 0.7% behind the most accurate entry. The chief weakness is statistical: no confidence intervals, seed variance, or significance tests accompany WER improvements as small as 0.09 percentage points, and subsampling, attention, and self-conditioning choices are partly evaluated with coupled baselines. Synthetic data construction and internal leaderboard comparisons are only sketched.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 2
● low 3
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 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
- Strided depthwise convolutions in the first two Conformer blocks achieve 8x temporal subsampling with better WER (5.83% vs 5.92%) than the FastConformer convolutional front end at the same reduction rate.
- Muon with Polar Express orthogonalization outperforms AdamW on every public test set (5.83% vs 5.99% average WER) under identical hyperparameters.
- Replacing 1x1 convolutions with linear layers and enabling the EFFICIENT_ATTENTION backend together yield a 49% inference speedup (inverse RTFx 9229 to 13715) at unchanged WER.
- The final model lies on the Open ASR leaderboard speed-accuracy Pareto frontier (4 Sep 2026) with 5.83% average WER while being about twice as fast as the fastest competitor, and the -nc companion model is 0.7% behind the most accurate entry.
Strengths
Strong speed-accuracy Pareto result with public verification
Figure 1 places granite-speech-5.0-470m-turboctc on the Open ASR leaderboard Pareto frontier (51 models, 4 Sep 2026) at 5.83% avg WER with roughly twice the RTFx of the fastest competitor, and the leaderboard itself is third-party maintained and externally checkable.
Well-controlled component ablations
Table 2 isolates each architectural choice against the final configuration: convolutional front end (5.92% vs 5.83%), grouped attention (5.94%), and removing self-conditioning (6.00%) all degrade, with degradations visible on nearly every individual test set.
First (to my knowledge) reported ASR training with Muon + Polar Express, fairly compared
High reproducibility disclosure
Transparent AI-usage statement and honest speedup attribution
Weaknesses
No variance or significance analysis for small WER deltas
Some ablation baselines are not factor-isolated
Comparisons against a moving leaderboard snapshot with unclear significance
Synthetic data and distillation teacher under-specified
Internal test sets and long-form claims not substantiated in-text
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
This Interspeech-style system paper describes granite-speech-5.0-470m-turboctc, a 470M-parameter encoder-only CTC Conformer for English short-form ASR. The central claims are: (1) strided depthwise convolutions give better pyramidal subsampling than the FastConformer front end; (2) Muon with Polar Express orthogonalization beats AdamW for Conformer training; (3) two inference-level changes give a 49% RTFx speedup at equal WER; and (4) the resulting model sits on the Open ASR leaderboard speed-accuracy Pareto frontier (4 Sep 2026), roughly twice as fast as the fastest competitor. The companion -nc model is 0.7% behind the top-accuracy entry, a speech LLM. The scope is deliberately narrow: encoder-only, English, short-form, greedy decoding — the paper does not claim state-of-the-art accuracy, and says so.
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 is IBM Granite 5.0 Turbo CTC?
It is a 470-million-parameter, encoder-only automatic speech recognition model built on a Conformer architecture trained with CTC loss. It is designed for fast English short-form transcription and is available on Hugging Face (ibm-granite/granite-speech-5.0-470m-turboctc) under a permissive license.
How accurate is Granite 5.0 Turbo CTC compared with other ASR models?
On the public test sets of the Open ASR Leaderboard it achieves about 5.83% average word error rate after its first training stage (5.72% after full training), placing it on the leaderboard's speed-accuracy Pareto frontier as of 4 September 2026. The companion model trained with extra data (-nc variant) was reported 0.7% behind the most accurate entry, which was a speech LLM.
Why is the model so fast at inference?
Two engineering changes drive the speed: chunked (block-diagonal) attention computed with an efficient attention kernel instead of a math backend, and replacing 1x1 convolutions with mathematically equivalent linear layers. Together these raised the inverse real-time factor from 9,229 to 13,715 (a 49% speedup) on an H100 with batch size 128, with no change in accuracy.
What data was the model trained on?
Roughly 60,000 hours of publicly available English speech: MLS (44.6k hours), YODAS, CommonVoice-17, Librispeech, VoxPopuli, AMI, and Earnings-22, plus about 2,700 hours of synthetic data (concatenated multispeaker audio and TTS-generated number/URL/email utterances created with gpt-oss models and StyleTTS 2). No proprietary data is used for the main model.
What is novel about the training method?
The base model is trained with the Muon optimizer using Polar Express orthogonalization, which the authors report beats AdamW on every test set (5.83% vs 5.99% average WER) under matched hyperparameters. Training also uses knowledge distillation from the granite-4.1-8b-base LLM and robust fine-tuning with speed, noise, reverberation, and codec perturbations followed by parameter interpolation.
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.
- Figure 1 (Pareto scatter) and Figure 2 (block diagrams) are only available as captions in the extracted text; competitor identities and exact axis positions cannot be verified from the text alone.
- The extraction appears complete through the references, but supplementary details (e.g., full leaderboard numbers, -nc training data) may exist outside the paper.
- Code availability is not explicitly stated in the text; only model weights are said to be downloadable, so codeAvailable=false reflects the paper's statements, not the Hugging Face repo contents.
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.