[clkao]

Add an optional mlx-qwen3-asr adapter using the native incremental stream. Model loading and warmup happen once at startup; decoding runs in WLK's counted inference executor under a shared lock, with separate state and language/context options for each session.

Session language validation uses the model’s declared 30-language inventory; unsupported Whisper languages are rejected. Stable text is committed directly. Pauses, speaker changes and EOF drain finish_streaming() without a second full-utterance decode. Preserve spaces and reject revisions of an already confirmed prefix instead of duplicating text. The proposed generic ASR transform layer, timestamp wrapper and their large unit-test suite are removed. Timestamps describe chunk bounds, not word alignments.

Validation: The final head passes all required CI checks, including real-audio boundary scenarios; three adapter scenarios cover boundaries, session isolation and failed prefix revisions. A frozen extra installation on M5 completed real English/French/Chinese streams. Ruff and lock checks passed. The lock adds the candidate package without upgrading existing package versions.

The evaluated head (24082a0) includes explicit pipeline failure reporting and schema-3.1 packet pacing from #443. A real shared-model check also matches sequential and concurrent English/Chinese transcripts exactly.

The completed M5 evaluation does not support merging this head. All sixteen reports were measured on AC power (#444, benchmarks/evaluations/m5-20261107/). English/French WER is 27.65%/44.68% and Chinese CER 28.46%: 16.0/27.0/2.7 points worse than Whisper small and 6.3/20.0/16.6 points worse than Qwen Metal with the same model revision. EOF p95 is slower than both baselines in every pass, and the MLX allocator peak is higher. Most of the gap with Metal is insertions and deletions: hypotheses drop spans at segment boundaries and repeat others, which points to the adapter or incremental finalizer rather than the model.

The ten-minute Chinese stream fails reproducibly after 101.5 s: at a pause, start_silence() calls finish_streaming() with a 160-sample tail and mlx-qwen3-asr 0.3.5 raises Audio too short for log-mel extraction. The restored README, badge and published figures are unchanged.

[QuentinFuxa]

Sorry for the long silence on this one. I measured it against the existing backends on an M5 (the full series is in #444) and ran it on real audio, and it is not mergeable yet. The integration itself is clean: the adapter is small, the only conflict with main is uv.lock, and the tests pass. The problem is the output.

Measurements (#444, 90 FLEURS clips per language, three passes)

	Native Qwen (this PR)	Qwen Metal, same weights	Whisper small MLX
English WER	27.7%	21.4%	11.7%
French WER	44.7%	24.7%	17.7%
Chinese CER	28.5%	11.8%	25.8%

EOF p95 is also slower than both baselines (0.9 to 1.9 s). Since Metal uses the same model revision, the gap comes from the adapter or the incremental finalizer, not the model. In French it is mostly insertions and deletions: 441 and 330, against 81 and 84 for Metal.

What I see on real audio

On audio_tests/00_00_16_french_1_speaker.wav, "là il y a eu tel second de silence et je" comes out twice in a row. On 00_00_30_english_3_speakers.wav, "It does a pretty good job filtering noise." is repeated, the next sentence is cut and the end of the file is missing.

I traced the repetition to finish_streaming() in mlx-qwen3-asr 0.3.5. The state text is correct before the call and doubled after it. When the tail decode adds nothing, the library re-transcribes the last window (last chunk plus pending tail) and appends the result to the existing text, so the same words appear twice. That refinement call also does not receive the session language. The adapter calls finish_streaming() at every pause and speaker change, so this happens at every segment boundary.

Two other things:

• The ten-minute Chinese stream fails after 101.5 s, reproducibly. At a pause, start_silence() calls finish_streaming() with a 160-sample tail and the library raises Audio too short for log-mel extraction.
• In my trace stable_text stays empty for the whole utterance, so nothing is committed before the pause. With 2 s chunks and the first two chunks kept unstable, short utterances never reach a stable prefix. That explains the latency numbers.

What would make it mergeable

1. No repeated text at boundaries: either disable the tail refinement (enable_tail_refine) or deduplicate in the adapter, with a test that goes through AudioProcessor on a recording with pauses.
2. No exception on a short tail at a pause.
3. A rerun of the three FLEURS languages showing it is at least on par with Qwen Metal, since the weights are the same.

The first one may be better fixed upstream in mlx-qwen3-asr. If you don't have time for this now, no problem, tell me and I'll mark the PR as on hold rather than leave it open without an answer.

[clkao]

d538a8e fixes all three. It sits on top of this branch, so no existing commit was rewritten, and I
merged main so the uv.lock conflict is gone.

• Duplication at boundaries. The tail refinement pass you traced is gone in 0.4.2 and later
  (PR #26). This branch requires 0.4.4, and the
  adapter commits only the text that clients do not already hold. test_pause_in_audio_commits_each_sentence_once
  drives AudioProcessor over audio with pauses.
• Short tail at a pause. Version 0.4.4 still rejects input below 200 samples, so the adapter
  keeps the decode window at or above 400 samples and ends the segment instead of raising. The
  600 s Chinese stream from #444 now runs to the end: 1 of 1 ok, 240 tokens, 8 lines, EOF 0.001 s.
• Quality, same weights. Model pinned to snapshot_revision 5eb1441… with a sha256 for each
  file. Corpus corpus_sha256 95586eb7…. Command: wlk bench --warmup --speed 1, 30 clips per
  language.

	this branch	qwen-native before	Qwen Metal	Whisper small
zh CER	11.15%	28.46%	11.85%	25.77%
en WER	13.89%	27.65%	21.36%	11.66%
fr WER	14.39%	44.68%	24.68%	17.66%

Likely cause of the gain: this backend commits text after a re-decode of a window of up to 30 s
that includes the pending tail. The Metal path commits with 0.64 s of right context
(qwen3_streaming_right_context_ms), and that horizon falls inside a word in English and French,
but rarely in Chinese. Two facts fit that: the two systems give byte-identical hypotheses on
16 of 30 Chinese clips and 3 of 30 English clips, and substitutions carry the whole gap while
insertions are equal.

[QuentinFuxa]

Big improvement, thanks. I checked it on the same manifest (one paced pass, --speed 1, French and English): English WER is 13.89%, exactly your figure, and French is 14.97% (you have 14.39%, close enough for a single pass). EOF p95 is 0.19 s, all 60 clips ok, valid timestamps. The ten-minute crash and the English repetitions are gone on my recordings too.

One bug is left before I can merge: revisions near the start of a segment still duplicate text. On audio_tests/00_00_16_french_1_speaker.wav the output reads "là il y a eu telLà il y a eu tel second de silence". In the FLEURS run, 7 of the 30 French clips have the same pattern, for example fleurs_fr_1935: "cependant ce n'est pas vrai bienCependant ce n'est pas vrai bien qu'il y ait…", and fleurs_fr_1975: "c'est le plusC'est le plus petit pays…".

The trace on the French recording:

source_before = 'là il y a eu tel'
candidate     = 'Là il y a eu tel second de silence et je repars'
delta         = 'Là il y a eu tel second de silence et je repars'

The model only capitalized the first word. _append_only_delta sees that the candidate does not start with the source and that it is longer, so it takes the part after the common prefix. The common prefix is empty because of the capital L, so the whole segment is appended again, and with no separator. Any revision of an early word does the same.

Since clients cannot take back text, I'd suggest appending only what goes beyond what they already hold: when the candidate rewrites the source and is at least as long, emit candidate[len(source):] (adjusted to a word boundary) and accept that the earlier wording stays. A case- and whitespace-insensitive comparison would already catch the capitalization cases. A regression test with exactly this trace would pin it.

With that fixed I'm happy to merge: French beats both baselines, Chinese matches Metal by your numbers, and English is close to Whisper small.

[clkao]

Fixed in db10f6c. Your trace now yields delta = ' second de silence et je repars'.

The cause was the case-sensitive shared-prefix comparison. Your candidate only capitalises the
first word, so the shared prefix measured zero and the segment went out again. Comparing with
casefold makes it a candidate that extends the committed text.

One deviation from your suggestion: I did not slice at len(source). When a rewrite cuts inside a
word, slicing at len(source) drops the first letters of that word, and it broke the four upstream
#31 cases in the test file. Keeping the slice at the shared prefix, with a case-insensitive
comparison, fixes your trace and preserves those. For the cut-inside-a-word case the test asserts
that what clients hold plus the delta reconstructs the candidate, so nothing is sent twice and no
word is lost.

Tests replay your exact trace, the fleurs_fr_1935 form of it, and that cut-inside-a-word case.

What I could not do: reproduce the 7 of 30 on French clips. audio_tests/ is not in the repo, and
in my FLEURS runs no French hypothesis repeats its opening span once legitimate repeats are
excluded ("ce n'est pas vrai … ce n'est pas une carte"). French weighted WER moved from 14.39% to
13.92% with the fix, so it did change output. Please re-run your check on
00_00_16_french_1_speaker.wav and on those French clips.