theres something very wrong in the speech to text #16800

Activity

1. niksedk commented on Nov 10, 2026

   

   niksedk

   on Nov 10, 2026

   Member

   More actions

   Thanks. Your two runs are not quite the same setup, so a few questions and things to try:

   1. Different model. WhisperJAV used neosophie/Qwen3-ASR-1.7B-JA. SE's Japanese Qwen3 model is a different fine-tune (qwen3-asr-1.7b-ja-anime).
   2. Vocal separation before VAD. Your WhisperJAV command runs HTDemucs before FireRedVAD (--pass1-enhance-for-vad). In SE, please turn on Isolate speech. It does the same job (with Mel-Band RoFormer), so the VAD doesn't cut on music and sound effects.
   3. Forced aligner. Qwen3 doesn't make timestamps on its own, so the timing comes from the aligner. Which one was selected? Please try Qwen3 forced aligner (or the Japanese wav2vec2 aligner).

   New Japanese model: #16802 (merged) adds Haas-ja v1.0 to Crisp ASR Parakeet. It is a Japanese-only model with the best open CER on asr-ja-bench, and as a Parakeet (TDT) model it makes its own timestamps. It will be in tomorrow's beta - please give it a try (language: Japanese, with Isolate speech + FireRedVAD).

   If it's still off, please tell us:

   • Your SE version (Help → About).
   • What "way off" means: too early, too late, too long, or several sentences merged into one cue?
   • The SE log from the Speech to text window. The full CrispASR command line is in there.

   A short sample (1–2 minutes) with both SRT files would help a lot, if you can share one.
2. gojetablade commented on Nov 11, 2026

   

   gojetablade

   on Nov 11, 2026

   Author

   More actions

   Thanks. Your two runs are not quite the same setup, so a few questions and things to try:

   1. Different model. WhisperJAV used neosophie/Qwen3-ASR-1.7B-JA. SE's Japanese Qwen3 model is a different fine-tune (qwen3-asr-1.7b-ja-anime).
   2. Vocal separation before VAD. Your WhisperJAV command runs HTDemucs before FireRedVAD (--pass1-enhance-for-vad). In SE, please turn on Isolate speech. It does the same job (with Mel-Band RoFormer), so the VAD doesn't cut on music and sound effects.
   3. Forced aligner. Qwen3 doesn't make timestamps on its own, so the timing comes from the aligner. Which one was selected? Please try Qwen3 forced aligner (or the Japanese wav2vec2 aligner).

   New Japanese model: #16802 (merged) adds Haas-ja v1.0 to Crisp ASR Parakeet. It is a Japanese-only model with the best open CER on asr-ja-bench, and as a Parakeet (TDT) model it makes its own timestamps. It will be in tomorrow's beta - please give it a try (language: Japanese, with Isolate speech + FireRedVAD).

   If it's still off, please tell us:

   • Your SE version (Help → About).
   • What "way off" means: too early, too late, too long, or several sentences merged into one cue?
   • The SE log from the Speech to text window. The full CrispASR command line is in there.

   A short sample (1–2 minutes) with both SRT files would help a lot, if you can share one.

   I'll test it
3. gojetablade commented on Nov 12, 2026

   

   gojetablade

   on Nov 12, 2026

   Author

   More actions

   Thanks. Your two runs are not quite the same setup, so a few questions and things to try:

   1. Different model. WhisperJAV used neosophie/Qwen3-ASR-1.7B-JA. SE's Japanese Qwen3 model is a different fine-tune (qwen3-asr-1.7b-ja-anime).
   2. Vocal separation before VAD. Your WhisperJAV command runs HTDemucs before FireRedVAD (--pass1-enhance-for-vad). In SE, please turn on Isolate speech. It does the same job (with Mel-Band RoFormer), so the VAD doesn't cut on music and sound effects.
   3. Forced aligner. Qwen3 doesn't make timestamps on its own, so the timing comes from the aligner. Which one was selected? Please try Qwen3 forced aligner (or the Japanese wav2vec2 aligner).

   New Japanese model: #16802 (merged) adds Haas-ja v1.0 to Crisp ASR Parakeet. It is a Japanese-only model with the best open CER on asr-ja-bench, and as a Parakeet (TDT) model it makes its own timestamps. It will be in tomorrow's beta - please give it a try (language: Japanese, with Isolate speech + FireRedVAD).

   If it's still off, please tell us:

   • Your SE version (Help → About).
   • What "way off" means: too early, too late, too long, or several sentences merged into one cue?
   • The SE log from the Speech to text window. The full CrispASR command line is in there.

   A short sample (1–2 minutes) with both SRT files would help a lot, if you can share one.

   its a tiny bit better -- but it still seems to be way off on the timing even with audio that is perfectly hearable -- no background noise or loud music or other people talking
   you can do your own testing if you want with this test file Ive been using
   https://www.mediafire.com/file/lfzifscv5m0mjtp/file.zip/file
4. gojetablade commented on Nov 13, 2026

   

   gojetablade

   on Nov 13, 2026

   Author

   More actions

   so I did some more testing and so far the best test came out like this

   whisperjav.main "S:\TEMP FOLDER\file\file.mp4" --ensemble --ensemble-serial --pass1-pipeline xxl --xxl-exe "C:\Users\user\AppData\Roaming\Subtitle Edit\SpeechToText\Purfview-Faster-Whisper-XXL\faster-whisper-xxl.exe" --pass1-model large-v3 --pass2-pipeline qwen --pass2-sensitivity balanced --pass2-qwen-params "{"framer": "vad-grouped"}" --pass2-scene-detector semantic --pass2-speech-segmenter firered-vad --pass2-speech-enhancer htdemucs --pass2-enhance-for-vad --pass2-model neosophie/Qwen3-ASR-1.7B-JA --merge-strategy smart_merge --output-dir source --subs-language native --language japanese --no-asr-telemetry --model-refresh-audio-minutes 20 --log-file "S:\TEMP FOLDER\file\raw_subs\whisperjav_run.log"

   a 2 pass run seemed to catch more stuff and fixing the timing pretty well
5. niksedk

   added a commit that references this issue on Nov 13, 2026

   Merge pull request #16882 from SubtitleEdit/claude/qwen3-vad-default-…

   ...

   9de67c9
6. timothybrush

   added a commit that references this issue on Nov 13, 2026

   Turn on VAD for Crisp ASR Qwen3, FireRedVAD on Auto (SubtitleEdit#16800)

   ...

   e30524d