Labyricorn/voicebox
1
0
Fork 0
mirror of https://github.com/jamiepine/voicebox.git synced 2026-10-03 17:15:19 -07:00
Code Issues Packages Projects Releases Wiki Activity

fix(backend): pass language/task directly to Whisper generate

Browse Source
This commit is contained in:
nox
2026-10-04 00:01:34 +00:00
committed by capy-ai-staging[bot]
parent e61c85365b
commit 7575a65e7f
1 changed files with 9 additions and 7 deletions
Download Patch File Download Diff File Expand all files Collapse all files
backend/backends/pytorch_backend.py
+9 -7
View File
@@ -358,15 +358,17 @@ class PyTorchSTTBackend:
) )
inputs = inputs.to(self.device) inputs = inputs.to(self.device)
# Generate transcription # Generate transcription.
# If language is provided, force it; otherwise let Whisper auto-detect # If language is provided, force it; otherwise let Whisper
# auto-detect. Pass language/task directly to generate() instead
# of building forced_decoder_ids — get_decoder_prompt_ids defaults
# to no_timestamps=True, which injects <|notimestamps|> and
# disables the timestamp tokens that return_timestamps=True (and
# therefore long-form decoding) depend on.
generate_kwargs = {} generate_kwargs = {}
if language: if language:
forced_decoder_ids = self.processor.get_decoder_prompt_ids( generate_kwargs["language"] = language
language=language, generate_kwargs["task"] = "transcribe"
task="transcribe",
)
generate_kwargs["forced_decoder_ids"] = forced_decoder_ids
with torch.no_grad(): with torch.no_grad():
predicted_ids = self.model.generate( predicted_ids = self.model.generate(