Breeze-ASR-26 — CTranslate2 / faster-whisper (INT8)
The fastest edge build: RTF 0.23–0.24 on CPU, real-time with headroom. Use this on servers, 8 GB+ hosts, or GPU. For 4 GB hosts or desktop apps use the GGML repo; for mobile/WASM use ONNX.
Part of the Breeze-ASR-26 edge family — the same MediaTek model in every runtime, pick by your constraint:
Repo Runtime RSS RTF (CPU 4-thread) Best for Breeze-ASR-26-ct2 CTranslate2 / faster-whisper ~2.9 GB 0.23–0.24 servers, 8 GB+ hosts, GPU Breeze-ASR-26-GGML whisper.cpp / MacWhisper 1.85 GB 0.40 (measured 2026-07-17) 4 GB hosts, desktop apps Breeze-ASR-26-ONNX sherpa-onnx / onnxruntime — 1.62–1.95 Android / iOS / WASM All Apache-2.0, derived from MediaTek-Research/Breeze-ASR-26. Measured on real multi-speaker Mandarin meeting audio. Mandarin does not regress; Taigi is transcribed as Mandarin meaning (not verbatim Taigi characters).
Usage
from faster_whisper import WhisperModel
m = WhisperModel("weemed/Breeze-ASR-26-ct2", device="cpu", compute_type="int8", cpu_threads=4)
segments, _ = m.transcribe("meeting.wav", language="zh", beam_size=1)
print("".join(s.text for s in segments))
For audio longer than 30 s, transcribe it in ≤30 s clips and concatenate the transcripts. A single call over the whole file is measurably slower (see below).
INT8 vs FP32: CER 4.69% (function-word/segmentation jitter only), 2.4x faster, quarter the size.
Re-measurement (2026-10-29)
A reply on discussion #3 suggested a different measurement method. We re-ran the benchmark on a new corpus and fixed two flaws in the July setup. The numbers above are the current ones.
Corpus. 300 s of a public Legislative Yuan (Taiwan) meeting recording — multi-speaker Mandarin, spontaneous speech, crosstalk, 16 kHz mono, 75.5 % speech density. The July audio file was never archived, so these figures come from a different recording; they land in the same range.
Hardware. Intel Core i7-13700KF, 4 threads, CPU only, OMP_NUM_THREADS=4.
| Build | Method | RTF |
|---|---|---|
| CT2 | ≤30 s clips, per-clip decode, timestamps on | 0.2412 |
| CT2 | ≤30 s clips, per-clip decode, timestamps off | 0.2257 |
| CT2 | one call over the whole file, timestamps on | 0.3602 |
| CT2 | one call over the whole file, timestamps off | 0.2670 |
| ONNX | 30 s windows (sherpa-onnx truncates the tail) | 1.6212 |
| ONNX | 29 s windows (no truncation) | 1.9479 |
What the reply changed, and what it did not:
- Splitting long audio into ≤30 s clips and combining transcripts is the win. One call over the same 300 s was 18 % (timestamps off) to 49 % (timestamps on) slower. We already did this; now it is measured.
return_timestamps=Falseis the OpenAI-whisper spelling. faster-whisper'stranscribe()has no such argument — the equivalent iswithout_timestamps=True. It saved 6.4 % RTF, but it is not content-neutral: in one of ten 30 s windows it emitted 10 characters where the timestamped run emitted 136. Some of the speed came from text that was never transcribed, so measure the transcript, not only the RTF.- Short-clip benchmarks do mislead. A 2.5 s slice measured RTF 1.30 — not real time, because Whisper pads every input to a 30 s window — and a repetitive 30 s clip measured RTF 2.28–2.83, because the decoder loops on repetitive audio.
- sherpa-onnx accepts strictly less than 30 s and truncates a 30.000 s window with only a log warning. Use 29 s windows. The July ONNX figure was measured that way.
The same audio also gave CT2 and ONNX a head-to-head transcript comparison: normalized similarity 0.8475 in favour of CT2, which drops fewer characters.
Corpus and measurement recipe
- Source. Legislative Yuan IVOD recording 170217 — a 410 s clip of 王義川 in the
外交及國防委員會 meeting of 2026-07-01 (731 speaker turns). Public open data:
https://ivod.ly.gov.tw/Play/Clip/1M/170217. The HLS host rejects requests without a
browser referer, so fetch it with
ffmpeg -nostdin -headers "Referer: https://ivod.ly.gov.tw/" -user_agent "Mozilla/5.0" -i <m3u8> -c copy source.mp4. - Cut. Decode the whole recording first, then take the densest 300 s:
ffmpeg -i source.mp4 -vn -ac 1 -ar 16000 -c:a pcm_s16le full.wavand thenffmpeg -ss 29.32 -t 300 -i full.wav -ac 1 -ar 16000 -c:a pcm_s16le corpus.wav. Seeking inside the mp4 and cutting the decoded wav are not the same bytes. - Corpus SHA-256:
670156c9156bd4aa48ac031523968f4a6a901987533f1f38f8b60749da18a56d(ffmpeg 4.4.2-0ubuntu0.22.04.1). A fresh CDN delivery re-encodes the container, so it lands within a few AAC samples of that digest rather than exactly on it. - Decode. Transcribe the corpus in ≤30 s clips, concatenate the transcripts, and time against the audio duration.
The July 2026 figures were measured on a corpus that was never archived. The figures in this card come from the recording described above.
Code-switched speech: a CER measurement (2026-10-30)
The benchmark above measures speed and transcript similarity. This card carried no character error rate for Mandarin-Taiwanese code-switching inside a single utterance, because no labelled sample existed. We built one and measured it.
The sample. 13 utterances from public Legislative Yuan (Taiwan) IVOD recordings, 2024-03 to 2026-07, 10 speakers. Each one is a gazette turn in which a Mandarin sentence carries Taiwanese vocabulary — 謝龍介「檢察官『會呼雞袂歕火』抓了一大堆」, 楊瓊瓔「這個不能斷電,要是斷電就『代誌大條了』」, 吳思瑤「過去的國際教育專法推了四年多, 『袂振袂動』」. The reference is the official gazette text, every character carries a language label, and the Taiwanese spans were reviewed by hand. 624 scored characters (39 Taiwanese, 585 Mandarin) over 149.3 s of audio.
The result — CT2 INT8, 4 threads, CPU:
| CER | miss rate (deletions + substitutions) | |
|---|---|---|
| whole utterance | 52.24 % (micro, 326 errors / 624 characters) | — |
| Mandarin segment | 51.28 % | 36.92 % |
| Taiwanese segment | 66.67 % | 53.85 % |
| IVOD | Turn | Speaker | Date | Scored chars | CER | Mandarin CER | Taiwanese CER |
|---|---|---|---|---|---|---|---|
| 150577 | 26 | 謝龍介 | 2024-03-28 | 166 | 48.2 % | 48.1 % | 50.0 % |
| 150911 | 30 | 李柏毅 | 2024-04-08 | 16 | 87.5 % | 60.0 % | 500.0 % |
| 150930 | 3 | 王美惠 | 2024-04-08 | 65 | 90.8 % | 90.0 % | 100.0 % |
| 151061 | 32 | 楊瓊瓔 | 2024-04-11 | 16 | 81.2 % | 76.9 % | 100.0 % |
| 151697 | 61 | 洪孟楷 | 2024-04-24 | 74 | 35.1 % | 36.1 % | 0.0 % |
| 153044 | 19 | 吳思瑤 | 2024-05-27 | 43 | 16.3 % | 12.8 % | 50.0 % |
| 153431 | 8 | 吳思瑤 | 2024-05-31 | 35 | 28.6 % | 20.0 % | 50.0 % |
| 155739 | 13 | 洪孟楷 | 2024-10-17 | 29 | 93.1 % | 96.3 % | 50.0 % |
| 157335 | 54 | 賴士葆 | 2024-11-21 | 16 | 68.8 % | 68.8 % | — |
| 157598 | 99 | 陳玉珍 | 2024-11-28 | 94 | 27.7 % | 28.0 % | 0.0 % |
| 159641 | 45 | 蔡易餘 | 2025-03-26 | 23 | 47.8 % | 45.5 % | 100.0 % |
| 170217 | 59 | 王義川 | 2026-07-01 | 30 | 76.7 % | 81.5 % | 33.3 % |
| 170730 | 62 | 賴士葆 | 2026-07-29 | 17 | 111.8 % | 118.8 % | 0.0 % |
Read the numbers with their limits. This is the hardest subset: the sample holds only the utterances where Taiwanese vocabulary was marked. It is not the CER of ordinary code-switched speech. The reference is the written gazette text, so a difference between the written sentence and the spoken words counts as an error. That hits the Mandarin segment too — its CER is 51.28 %, so most of the number is not the code-switching. 13 utterances and 624 characters is a small sample with a wide interval. The per-utterance CER runs from 16.3 % to 111.8 %. Placing each utterance on the audio is itself an error source. The gazette turn times of the IVOD API drift away from the audio. On one clip they are 10.3 s early, so a cut made from them holds a different sentence. These windows were placed with MMS-1B CTC, an independent model that reads characters and their times out of the audio. With an oracle window chosen by the evaluated model, the micro CER would be 45.19 %; that oracle is a bound, not a score. One thing a character CER cannot show: the model rewrites Taiwanese idioms into Mandarin ones. 吳思瑤's「袂振袂動」came back as「動彈不得」 — right in meaning, wrong in characters, and a character CER understates the real gap here.
Reproducing it. The sample ships with this card as
code-switch-cer-sample.json.
For every utterance it gives the IVOD clip, the gazette turn, the reference text, the
Taiwanese spans with their character offsets, and the window in seconds. Fetch the clip
from https://ly.govapi.tw/v2/ivods/<ivod_id>, decode the whole clip to 16 kHz mono, cut
window_start_s to window_end_s, transcribe, and compare against the reference
character by character.
This card and discussion #3 carry the method, the numbers and their limits.
- Downloads last month
- 144
Model tree for weemed/Breeze-ASR-26-ct2
Base model
openai/whisper-large-v2