chore: revert audio_transcription_service temperature change (out of scope)
Whisper temperature tuning belongs with the transcription stack work, not with WhatsApp call wiring. Restoring develop's value to keep the PR focused.
This commit is contained in:
@@ -67,9 +67,7 @@ class Messages::AudioTranscriptionService< Llm::LegacyBaseOpenAiService
|
||||
parameters: {
|
||||
model: WHISPER_MODEL,
|
||||
file: file,
|
||||
# 0.0 minimises Whisper's hallucinations on silence / sub-vocal segments —
|
||||
# higher temperatures produce phantom phrases that look like real transcripts.
|
||||
temperature: 0.0
|
||||
temperature: 0.4
|
||||
}
|
||||
)
|
||||
transcribed_text = response['text']
|
||||
|
||||
Reference in New Issue
Block a user