[Arthur031221]

Summary

funasr FILE -f tsv writes 0.000 as both start and end of its only row when the result has no sentence-level timestamps (an empty or missing sentence_info). -f srt already handles that case by spanning the token timestamps, or the audio length if there are none. This change moves that span logic into one helper, _fallback_bounds_ms, and uses it for both formats. SRT output is unchanged.

Type of change

• Bug fix

Validation

• python -m compileall funasr/cli.py tests/test_cli.py: passes.
• pytest tests/test_cli.py: 29 passed. With funasr/cli.py restored to the base commit, the new TSV test fails (0.000 0.000 hello world instead of 0.300 1.700 hello world) and the other 28 pass.
• The new SRT test passes on both base and branch; it pins the unchanged SRT behaviour.
• The tests use a stub AutoModel returning sentence_info: [] and token timestamps [[300, 800], [900, 1700]]. I did not run a real model.

User impact

Anyone running funasr audio.wav -m fun-asr-nano -f tsv got the whole transcript on one row timed 0.000 to 0.000. The fun-asr-nano entry in MODEL_CONFIGS has no punctuation model and the CLI does not add one for it, so with token timestamps present inference_with_vad sets sentence_info to [] (the "punc_model is required for sentence_timestamp" warning branch). I confirmed that branch by reading auto_model.py and by driving inference_with_vad with stubs, not with the real model.

Notes for reviewers

• The fix is in the CLI formatter, matching how the SRT fallback already handles this case. It does not add per-sentence TSV rows for such models: the output is still one row, now spanning the first to the last token timestamp.
• When there are neither sentence segments nor usable token timestamps (including an empty transcript), the row spans 0 to the audio duration, or 0 to 0 if the duration cannot be read. SRT already does the same.

[LauraGPT]

Reviewed e3a2821d489f5aa646139dfafd3b8ff415ce1cfb against its parent 4a84879ae8ca4d2430953b11c49d609217d93988. No actionable correctness findings in this change.

Independent CPU validation using the same complete tests/test_cli.py from this PR:

• Parent: 28 passed, 1 failed. The new TSV regression reproduces the incorrect 0.000 -> 0.000 bounds.
• PR head: 29 passed, 0 skipped/errors.
• Additional 14 boundary cases (26 assertions) passed: token bounds, dictionary/unsorted timestamps, actual WAV-duration fallback, unreadable audio, empty/unusable timestamps, unchanged sentence-segment output, and byte-equivalent SRT output.

The shared fallback preserves the existing SRT behavior while giving the single TSV row the available time span. Validation used model doubles at the existing test boundary; no real-model inference or full-repository test claim.

Commit metadata note: this head is unsigned and has no Signed-off-by trailer. This is a comment-only technical review, not a merge-policy approval.