Qwen3-ASR 0.6B for Core ML
This repository contains a Core ML conversion of
Qwen3-ASR 0.6B by the Qwen team, Alibaba Cloud,
at revision 7f1569a48a89f3e3f4dc3a5c9d28bddd903bc76c.
AnySpeechRecognizer runs it on Apple silicon
with the same API as Apple's SpeechTranscriber.
Languages
The model transcribes speech in 30 languages: Arabic, Cantonese, Chinese, Czech, Danish, Dutch, English, Filipino, Finnish, French, German, Greek, Hindi, Hungarian, Indonesian, Italian, Japanese, Korean, Macedonian, Malay, Persian, Polish, Portuguese, Romanian, Russian, Spanish, Swedish, Thai, Turkish, Vietnamese. It identifies the language of each segment, or the caller can set the language. The original model card also lists 22 Chinese dialects. This conversion was evaluated on Mandarin, Japanese, Korean, Cantonese, and English only.
Use
let transcriber = SpeechTranscriber(
model: QwenSpeechModel(variant: .small),
transcriptionOptions: [],
reportingOptions: [],
attributeOptions: []
)
A version of AnySpeechRecognizer that pins this repository downloads and verifies these files
through AssetInventory. Other versions can load a local copy of this repository with
QwenSpeechModel(variant: .small, assets: .directory(url)).
Limitations
- The model transcribes files. It does not stream, and it gives no volatile (partial) results. Each final result covers one segment of up to 40 seconds of audio.
- Results have no word timings, only the time range of their segment. The separate Qwen3-ForcedAligner-0.6B model, which predicts word timings, is not included.
- Results have no confidence values or alternatives, and the context (hotword) prompt of the original model is not supported.
- The language that the model identifies is not reported.
- The decoder runs on the GPU or the CPU. On macOS 27.0.1, Core ML cannot build it for the Neural Engine.
- This conversion was tested on a Mac with an M4 Max and macOS 27.0.1 only. It has not been tested on an iPhone or iPad.
Size and resources
| Item | Value |
|---|---|
| Files | 0.94 GB |
| Peak memory while transcribing on a Mac | 1.74 to 1.76 GB |
| Real-time factor (processing time divided by audio duration) | 0.22 to 0.32 |
| Core ML specification | iOS 18 and macOS 15 (stateful ML programs) |
The real-time factor was measured on a Mac with an M4 Max while another speech job used the GPU, so it is higher than on an idle Mac.
Evaluation
The table gives the error rates of AnySpeechRecognizer's runtime with this conversion
in file mode, measured on 2026-10-17 with its Tools/natural-speech-eval.py.
The FLEURS sets are the first 100 test files of each language.
Character error rates (CER) count characters after NFKC normalization and lowercasing,
without punctuation, symbols, or spaces.
For Cantonese, both texts are first converted to simplified characters with OpenCC,
because the references use traditional characters and the model writes either script.
Word error rates (WER) count words after lowercasing and removing punctuation.
The transformers column is the original model in transformers with greedy decoding
on the same files.
| Test set | Measure | Error rate | Substitutions / deletions / insertions | Transformers, same files | Real-time factor |
|---|---|---|---|---|---|
| FLEURS Mandarin (100 files) | CER | 7.11% | 85 / 133 / 31 of 3,502 | 7.14% | 0.22 |
| FLEURS Japanese (100 files) | CER | 9.49% | 321 / 84 / 74 of 5,048 | 9.31% | 0.23 |
| FLEURS Korean (100 files) | CER | 5.29% | 120 / 105 / 9 of 4,426 | 4.97% | 0.26 |
| FLEURS Cantonese (100 files) | CER | 8.15% | 234 / 23 / 30 of 3,522 | not run | 0.24 |
| LibriSpeech test-clean (252 files) | WER | 2.14% | 100 / 10 / 5 of 5,369 | 2.10% | 0.30 |
| LibriSpeech test-other (281 files) | WER | 4.65% | 181 / 41 / 19 of 5,178 | 4.63% | 0.32 |
The FLEURS references write numbers as digits, names in Latin script, and some English glosses in parentheses that speakers did not always read. The model often writes numbers in words and names in the local script, so these error rates are higher than those in the original model card.
Conversion and verification
The models were converted with this command in Tools/QwenConversion:
uv run python convert.py --variant 0.6b --dtype float32
The tools were Python 3.12.12, PyTorch 2.8.0, transformers 5.20.0, and coremltools 9.0 on macOS 27.0.1. The KV cache holds 1,024 tokens, and the decoder takes 16 tokens in each prediction.
Tools/QwenConversion/verify.py compared the conversion with the transformers model.
It compared the audio embeddings for each token,
the decoder's greedy token after each prefix of the transformers output,
and complete greedy transcripts.
| Model | Clips | Encoder mean / lowest token cosine | Decoder tokens that match, teacher-forced | Transcripts identical to generate() |
|---|---|---|---|---|
| PyTorch wrappers (before Core ML) | 8 synthetic speech (say) |
1.00000 / 1.000 | 311 of 311 | 8 of 8 |
| Core ML, CPU_AND_GPU | 8 synthetic speech (say) |
0.99901 / 0.958 | 311 of 311 | 8 of 8 |
| Core ML, CPU_AND_GPU | 16 FLEURS and LibriSpeech | 0.99829 / 0.512 | 559 of 562 | 13 of 16 |
The three differing transcripts differ in near-tied words: in Japanese, "カタルーニャを好み" for "カタルーニャがお好み" and "カタルーニャと" for "カタルーニャ語と"; in Korean, "일만" for "1만"; and in English, commas for dashes.
Files
| File | Size |
|---|---|
AudioEncoder.mlmodelc |
187.2 MB |
LICENSE |
11.3 kB |
NOTICE |
1.8 kB |
TextDecoder.mlmodelc |
754.3 MB |
qwen3_asr_config.json |
5.5 kB |
vocab.json |
2.3 MB |
conversion.json lists the size and SHA-256 digest of every file except itself and this card.
License
The original model and this conversion are licensed under the
Apache License, Version 2.0. See LICENSE and NOTICE.
Qwen3-ASR 0.6B by the Qwen team, Alibaba Cloud (https://huggingface.co/Qwen/Qwen3-ASR-0.6B-hf), licensed under the Apache License, Version 2.0 (https://www.apache.org/licenses/LICENSE-2.0). Converted to Core ML for AnySpeechRecognizer: the model is split into an audio encoder and a text decoder with a KV cache, weights are stored in float16, and weights of large layers are quantized to int8.