Breeze-ASR-26 β€” GGML (whisper.cpp, Q4_0 / Q5_0)

The smallest-footprint build: peak RSS 1.85 GB (Q4_0), fits a 4 GB host, and loads directly in whisper.cpp desktop apps (MacWhisper Pro, superwhisper) β€” drag the .bin in, no code. Slower than CT2 (RTF 0.40) but the RAM floor is 1 GB lower.

(Files are whisper.cpp GGML format. whisper.cpp GGUF support is in progress upstream, ggml-org/whisper.cpp#3316; GGUF builds will be added when the toolchain lands.)

Part of the Breeze-ASR-26 edge family β€” 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 β€” 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

# whisper.cpp
whisper-cli -m breeze-q5_0.bin -f meeting.wav -l zh -t 4

MacWhisper Pro / superwhisper: download breeze-q5_0.bin, drag it in as a custom model.

Build Disk Peak RSS RTF
Q4_0 848 MB 1.85 GB 0.40
Q5_0 1.1 GB 2.03 GB 0.62

Re-measurement note (2026-10-29)

The CT2 and ONNX cards describe the corpus and the measurement in full, including how to rebuild the audio from a public Legislative Yuan recording and the expected SHA-256 of the corpus. See CT2 for the recipe.

The CT2 and ONNX builds were re-benchmarked on 2026-10-29 after the reply on discussion #3; their numbers moved slightly (CT2 0.23–0.24, ONNX 1.62–1.95). These GGML figures are still the 2026-07-17 measurement β€” the whisper.cpp build was not part of that re-run, so treat 0.40 / 0.62 as older numbers on a different corpus, not as a re-measured value.

Code-switched speech: a CER measurement (2026-10-30)

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 on the CT2 build; the numbers are on the CT2 card and in the sample file that ships with it.

Short version: on 13 reviewed utterances of public Legislative Yuan IVOD recordings, each one a Mandarin sentence carrying Taiwanese vocabulary, the CT2 INT8 build scored 52.24 % CER overall (micro, 326 errors / 624 characters), 51.28 % on the Mandarin segment and 66.67 % on the Taiwanese segment. The sample is the hardest subset β€” only the utterances where Taiwanese vocabulary was marked β€” the reference is the written gazette text, and 13 utterances / 624 characters carry a wide interval. The windows were placed with an independent model (MMS-1B CTC), never with this build's own output.

We measured the CT2 build on that sample. This build was not run on it.

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support

Model tree for weemed/Breeze-ASR-26-GGML

Finetuned
(17)
this model