Skip to content

[qwen3_asr] Punctuation is stripped/filtered out in --text-out #430

Description

@shiyueGG

Description
When executing the qwen3_asr pipeline via audiocpp_cli.exe, the output generated by --text-out lacks all punctuation marks.

While the word-level timestamp JSON (--words-out) and the transcribed text are produced as expected, the punctuation is completely stripped out from the full text output.

Command Executed
Python
cmd = [
".\audiocpp_cli.exe",
"--task", "asr",
"--family", "qwen3_asr",
"--model", asr_model,
"--backend", "cuda",
"--audio", audio_file,
"--audio-chunk-mode", "vad",
"--session-option", f"qwen3_asr.vad_model_path={vad_model}",
"--session-option", f"qwen3_asr.forced_aligner_model_path={aligner_model}",
"--text-out", textout,
"--words-out", wordsout
]
Expected Behavior
Since the native Qwen3 ASR model naturally outputs transcriptions with punctuation marks, --text-out should preserve these punctuation marks in the final text output.

Actual Behavior
The text saved to --text-out appears to be filtered/stripped of all punctuation marks.

Impact
Without punctuation marks in the output text, it is difficult to split or align the transcript into standard subtitle segments (e.g., SRT/VTT), requiring an extra punctuation restoration step during post-processing.

Additionally, if unpunctuated text is strictly required for --text-out, users can easily extract or reconstruct it directly from the word-level data in --words-out.

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

Labels

No labels
No labels

Projects

No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions