fix(server): segment text-only OpenAI transcription responses

[LauraGPT]

Summary

Fixes the OpenAI-compatible /v1/audio/transcriptions fallback path for non-vLLM models such as sensevoice / paraformer when the underlying AutoModel.generate() result returns text but no sentence_info.

Before this patch, response_format=verbose_json returned segments: [] even when text was populated, which causes OpenAI-compatible clients that rely on segments for subtitles to render one large cue or lose timing entirely.

This patch:

• records fallback audio duration via soundfile.info()
• preserves model-provided sentence_info when available
• builds coarse OpenAI-style segments from long text-only results when sentence_info is missing
• keeps short text-only responses as one cue
• adds lightweight tests for the Kennedy-style long-text / empty-segments response shape

Validation

python3 -m pytest tests/test_server_app_openai_segments.py tests/test_fun_asr_nano_openai_response.py -q
# 4 passed

python3 -m py_compile funasr/bin/_server_app.py tests/test_server_app_openai_segments.py

git diff --check

[gemini-code-assist]

Code Review

This pull request introduces fallback segment generation for non-LLM models when only text is returned. It adds helper functions to split text into OpenAI-compatible segments based on character length and audio duration, and includes unit tests for this behavior. Feedback highlights two main issues: potential data loss when the audio duration is invalid or zero (where only the first segment of a split text is returned instead of the full text), and a redundant, dead-code check for an empty word list after the text has already been stripped and validated.

Important

The consumer version of Gemini Code Assist on GitHub is being sunset. Starting June 18, 2026, new organization installations will be blocked, and all code review activity will officially cease on July 17, 2026.
For more details on the timeline and next steps, please review the Help Documentation.

[gemini-code-assist]

When duration is less than or equal to 0 (which can happen if soundfile.info fails to parse the audio duration, e.g., for unsupported formats like MP3 on some platforms), and the transcribed text is long enough to be split into multiple parts, returning parts[0] will silently discard all subsequent parts of the transcription. This leads to severe data loss in the response. Instead, the entire text.strip() should be returned as a single segment when the duration is invalid or cannot be determined.

Suggested change

	if len(parts) == 1 or duration <= 0:
	return [{"start": 0.0, "end": max(float(duration), 0.0), "text": parts[0]}]
	if len(parts) == 1 or duration <= 0:
	return [{"start": 0.0, "end": max(float(duration), 0.0), "text": text.strip()}]

[gemini-code-assist]

Since text is stripped and checked for truthiness on lines 31-33, it is guaranteed to contain at least one non-whitespace character. Therefore, text.split() will never return an empty list, making the if not words: check dead code. We can safely remove this check to simplify the function.

Suggested change

	words = text.split()
	if not words:
	return [text[i : i + max_chars] for i in range(0, len(text), max_chars)]
	words = text.split()