Qwen3 ASR on your Mac, with word-level timestamps

The Qwen team at Alibaba released Qwen3 ASR on 29 January 2026 as three open models under the Apache 2.0 license: Qwen3-ASR-1.7B, Qwen3-ASR-0.6B and Qwen3-ForcedAligner-0.6B. TranscribeX runs Qwen ASR locally on your Mac.

Qwen3 ASR models

Open ASR figures for 1.7B from the leaderboard paper (Srivastav et al., arXiv 2510.06961, Table 3); for 0.6B from the Hugging Face evaluation panel on its model card. LibriSpeech and FLEURS figures are Qwen’s own, from the 1.7B and 0.6B model cards, and are not directly comparable with leaderboard numbers. RTFx is measured on datacenter GPUs. Checked October 2026.

Measured on a Mac

We ran Qwen3 ASR 0.6B (the 8-bit build TranscribeX uses) on the same 8-minute English test file as the other models. It scored a 4.8% error rate: better than Whisper Tiny, a little behind Whisper Large V3 and Parakeet V2 on English. Its strength is the languages the English-first models don’t cover.

Our own test, run October 17, 2026 on a Mac with an Apple M4 Pro chip and 24 GB of memory, macOS 27.0.1. Audio: 73 clean read-speech clips from the LibriSpeech validation set (hf-internal-testing/librispeech_asr_dummy on Hugging Face), joined into one 8-minute file, 1,150 reference words. Each model was run twice through the TranscribeX command line and the second run is shown; times are end to end, including loading the model. Word error rate uses simple normalization (lowercase, punctuation removed). Clean read speech is easier than a real meeting, so expect higher error rates on your own calls, and treat differences under one point as noise on a sample this size. Qwen3 ASR was not timed in this round; Whisper Large V3 Turbo, Parakeet V3 and Nemotron were not part of it.

The 30 languages are Chinese, English, Cantonese, Arabic, German, French, Spanish, Portuguese, Indonesian, Italian, Korean, Russian, Thai, Vietnamese, Japanese, Turkish, Hindi, Malay, Dutch, Swedish, Danish, Finnish, Polish, Czech, Filipino, Persian, Greek, Hungarian, Macedonian and Romanian. That is fewer than Whisper’s 99. The 22 Chinese dialects are the bigger difference: Whisper treats Chinese as one language, plus Cantonese since Large V3.

0.6B or 1.7B?

The 1.7B model is clearly more accurate, especially outside English: Qwen reports 4.90% WER on multilingual FLEURS for 1.7B against 7.57% for 0.6B. The 0.6B model is smaller and quicker, and on English it is close to Parakeet V3. In TranscribeX, Qwen3 ASR 0.6B is available as an 8-bit build of about 1 GB.

TranscribeX runs the 0.6B model, an 8-bit build of about 1 GB, which keeps it quick on any Apple Silicon Mac. If you transcribe Chinese, Japanese, Korean or Cantonese regularly, it is the model to try first. For English-only audio, Parakeet V2 was more accurate in our test.

The Qwen3 forced aligner

A speech recognition model is good at finding the words. It is less good at saying exactly when each word started. Forced alignment is a separate step that takes the transcript and the audio and pins every word to a precise time. That is what makes subtitles appear on the right syllable, lets the player highlight each word as it is spoken, and makes word-level exports usable for editing.

Qwen3-ForcedAligner-0.6B is an LLM-based, non-autoregressive timestamp model. Qwen reports an average alignment error of 42.9 ms on its test data, against 129.8 ms for the NeMo Forced Aligner and 133.2 ms for WhisperX. On human-labeled data the error was 32.4 ms. It handles up to five minutes of speech per pass in 11 languages: Chinese, English, Cantonese, French, German, Italian, Japanese, Korean, Portuguese, Russian and Spanish.

Alignment figures from the Qwen3-ASR repository and the ForcedAligner model card. They are Qwen’s own measurements.

In TranscribeX, the aligner runs as a companion to Qwen3 ASR to give precise word-level timestamps, which you can use in subtitle exports (SRT, VTT, bilingual subtitles), JSON exports and the synced player.

Trade-offs to know about

• Fewer languages than Whisper. For languages outside the 30, use Whisper Large V3.
• Not the fastest. On the leaderboard its RTFx is in the same range as Whisper Large V3 and far below Parakeet. If speed on English or European audio matters most, Parakeet wins.
• Streaming is not part of the local setup. Qwen’s own streaming mode depends on a vLLM server backend. For live captions, TranscribeX uses Nemotron.

Qwen ASR in TranscribeX

Qwen3 ASR is part of TranscribeX Pro, together with Whisper, Parakeet and Nemotron, speaker diarization and AI summaries with a local model. Everything runs on your Mac after a one-time model download, so it also works offline. MacWhisper offers Qwen3 ASR as well; here is how the apps differ.