fix(stt): strip Qwen3-ASR response prefix

[LauraGPT]

Summary

• strip the structured language ...<asr_text> prefix returned by Qwen3-ASR-compatible OpenAI transcription endpoints
• apply the normalization after preserving the existing string, SDK .text, dictionary, and fallback response handling
• add regression coverage for all three structured response shapes
• preserve the current provider-aware STT configuration architecture

Context

This refreshes the valid fix from #8773 onto the maintained main line. The original branch is now reported as conflicting and also contains an unrelated rollback of the current STT model configuration behavior.

The implementation retains attribution to @malaiwah through the original author trailer.

Validation

• rebased onto current main@56e2ba5e7; git range-diff confirms the one-commit patch is unchanged from prior head a6792ca4a
• 235 passed across managed media gateways, transcription providers, plugin dispatch, and gateway STT configuration tests on exact head 113a0aa25
• the three new response-shape tests were observed failing before the implementation and passing afterward
• ruff check passed for both changed files
• Python compilation and git diff --check passed
• an extra whole-file ruff format --check probe reproduces the same pre-existing formatting result on current main; the two-file PR patch remains unchanged

[teknium1]

Thanks for refreshing the Qwen3-ASR fix and preserving the current STT configuration work. The reported defect remains present: current main returns raw string transcription responses from tools/transcription_tools.py:1880-1881, and _transcribe_openai() forwards that value at tools/transcription_tools.py:1393.

Problems

• The new marker condition matches any <asr_text> occurrence, while the documented Qwen envelope is language ...<asr_text>. Since this helper also serves Mistral and ElevenLabs (tools/transcription_tools.py:1447, 1630), it can truncate an otherwise legitimate transcript containing the literal marker. The new tests cover only stripping cases.

Suggested changes

• Restrict the normalization to the start-anchored structured envelope and add a pass-through regression for a non-envelope string containing <asr_text>.

Automated hermes-sweeper review.

[LauraGPT]

Synced with current main and pushed exact head d59c2ca19ed9017ddbbf01adbf22c8c3602b713a; the branch is now 3 commits ahead / 0 behind and remains limited to tools/transcription_tools.py plus focused transcription tests.

Local validation on ind-gpu8 / Python 3.12.3:

• python3 -m pytest tests/tools/test_transcription_tools.py::TestExtractTranscriptText tests/tools/test_managed_media_gateways.py::test_extract_transcript_text_strips_qwen3_asr_prefix -q (5 passed)
• python3 -m pytest tests/tools/test_transcription_tools.py -q -k 'TranscribeOpenAI or TranscribeElevenLabs or TranscribeGroq or ExtractTranscriptText' (17 passed, 92 deselected)
• python3 -m pytest tests/tools/test_managed_media_gateways.py -q -k 'qwen or transcription or asr_text or conflict' (4 passed, 13 deselected)
• python3 -m compileall -q tools/transcription_tools.py tests/tools/test_transcription_tools.py tests/tools/test_managed_media_gateways.py
• git diff --check

A broader tests/tools/test_transcription_tools.py -k 'qwen or asr_text or transcript' run still reaches 106 passed and then fails only on the 3 Mistral tests because this host's system Python blocks lazy mistralai==2.4.8 installation with PEP 668 (externally-managed-environment). That failure is outside the Qwen3-ASR transcript-prefix diff and matches the previously reported local environment boundary.

[LauraGPT]

Addressed the sweeper review on the current PR head.

Change made:

• restrict Qwen3-ASR cleanup to a start-anchored language <code> [<audio_language>...</audio_language>] <asr_text> envelope
• preserve ordinary transcript text that merely contains <asr_text>, including text beginning with “Language ...”

Validation on ind-gpu8 / Python 3.12.3:

• python3 -m pytest tests/tools/test_transcription_tools.py::TestExtractTranscriptText tests/tools/test_managed_media_gateways.py::test_extract_transcript_text_strips_qwen3_asr_prefix -q -> 6 passed
• python3 -m pytest tests/tools/test_transcription_tools.py -q -k 'TranscribeOpenAI or TranscribeElevenLabs or TranscribeGroq or ExtractTranscriptText' -> 18 passed, 92 deselected\n- python3 -m pytest tests/tools/test_managed_media_gateways.py -q -k 'qwen or transcription or asr_text or conflict' -> 4 passed, 13 deselected\n- python3 -m compileall -q tools/transcription_tools.py tests/tools/test_transcription_tools.py tests/tools/test_managed_media_gateways.py\n- git diff --check\n

[LauraGPT]

Current-head verification for 8e028103780c after the Qwen3-ASR prefix cleanup:

python -m pytest tests/tools/test_transcription_tools.py -q -k 'qwen or Qwen or managed_media or transcription_prefix or language'
python -m pytest tests/tools/test_managed_media_gateways.py::test_extract_transcript_text_strips_qwen3_asr_prefix -q
python -m py_compile tools/transcription_tools.py tests/tools/test_transcription_tools.py tests/tools/test_managed_media_gateways.py
git diff --check origin/main...HEAD
git rev-list --left-right --count origin/main...HEAD

Results:

• focused transcription cleanup tests: 3 passed, 107 deselected;
• managed-media Qwen3-ASR prefix cases: 3 passed;
• compile and whitespace checks passed;
• branch is 0 behind / 4 ahead of origin/main.

I also tried the broader two-file pytest selection. It reaches 117 passing tests, then fails only in unrelated lazy-dependency paths because this host's system Python is PEP 668 externally managed and cannot lazy-install mistralai / fal-client; the failures occur before those tests' mocks can exercise provider behavior. The Qwen3-ASR prefix path touched by this PR is covered by the focused tests above.

[LauraGPT]

Synced this PR with current NousResearch/main after it had drifted behind. New head is 168b09ac8 (168b09ac8d9af39d5a8767df0117de8e7bfd479e). The merge was clean, and the final PR diff remains scoped to the Qwen3-ASR transcription envelope handling plus focused tests:

• tools/transcription_tools.py
• tests/tools/test_transcription_tools.py
• tests/tools/test_managed_media_gateways.py

Fresh validation on the synced head:

python3 -m pytest -q tests/tools/test_transcription_tools.py::TestExtractTranscriptText tests/tools/test_managed_media_gateways.py::test_extract_transcript_text_strips_qwen3_asr_prefix
# 6 passed in 0.20s

git diff --check
# passed

I also ran the broader transcription/media-gateway test files; the Qwen-related tests passed, while unrelated optional-provider tests hit this runner's system-Python PEP 668 guard while trying to lazy-install mistralai / fal-client.

[teknium1]

Merged into main via consolidated salvage PR #73510 (merge c0c5dac531). Your Qwen3-ASR language …<asr_text> envelope strip (3 commits, anchored regex) was cherry-picked with your authorship, with co-credit to @malaiwah (#8773, earliest report/fix).

Your contribution is credited to you in git history. Thank you! Closing this PR as merged-via-salvage.