[David-Wu1119]

Summary

funasr-server answers response_format=text on /v1/audio/transcriptions with JSONResponse(result["text"]), so the body is a JSON string literal ("say \"hi\"", application/json). OpenAI's transcription API returns the bare transcript for text, and the OpenAI Python SDK returns response.text directly for that format (openai.lib._parsing._audio.get_transcription_response_format_type maps text to str), so SDK and curl users receive the quotes and escape characters as part of the transcript.

client.audio.transcriptions.create(model="sensevoice", file=f, response_format="text")
# main:   '"say \\"hi\\""'
# branch: 'say "hi"'

Fix: return PlainTextResponse for this branch. json and verbose_json are unchanged.

Type of change

• Bug fix
• Documentation
• Example or demo
• Runtime or deployment
• Benchmark or evaluation
• Model/training change

Validation

Environment: macOS arm64, Python 3.11.15, CPU-only torch 2.15.0, fastapi 0.162.4 / starlette 1.10.0, openai 3.49.0, pytest 9.2.0.

• New test_text_response_format_returns_plain_text in tests/test_server_app_openai_segments.py (the stubbed fastapi.responses module gains a PlainTextResponse stub). It fails on main (the route returns the JSON-wrapped value) and passes with this change: python -m pytest -q tests/test_server_app_openai_segments.py gives 1 failed / 23 passed on main, 24 passed here.
• Neighbours: tests/test_service_docs_contract.py, tests/test_openai_api_n8n_compat.py, tests/test_fun_asr_nano_openai_response.py, tests/test_openai_api_docker_validation.py, tests/test_docs_funasr_install_commands.py pass. tests/test_gradio_app.py fails the same 19 cases on main and here because gradio is not installed in the test environment.
• End to end with real FastAPI: create_app(device="cpu", preload_model="sensevoice") with a stub AutoModel, POST /v1/audio/transcriptions (response_format=text) through fastapi.testclient.TestClient and through openai.OpenAI(..., http_client=TestClient). main: application/json, body "say \"hi\""; this branch: text/plain; charset=utf-8, body say "hi".
• python -m compileall funasr examples tests
• Docs or links checked
• Runtime/deployment command tested (in-process TestClient against create_app; no model weights, no GPU, no running funasr-server process)

User impact

OpenAI-compatible clients (OpenAI SDK, LiteLLM-style proxies, scripts migrating from Whisper) that request response_format=text get the transcript itself.

Notes for reviewers

srt and vtt still fall through to the json branch; that is out of scope here.

🤖 Generated with Claude Code

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has insufficient remaining quota. Try requesting a review at a lower effort level.

An organization owner can choose the repository organization as the billing source for reviews requested by members with a paid Copilot license. Learn about Copilot code review billing.