fix(transcription): drop Whisper temperature to 0 to suppress hallucinations
This commit is contained in:
@@ -67,7 +67,9 @@ class Messages::AudioTranscriptionService< Llm::LegacyBaseOpenAiService
|
||||
parameters: {
|
||||
model: WHISPER_MODEL,
|
||||
file: file,
|
||||
temperature: 0.4
|
||||
# 0.0 minimises Whisper's hallucinations on silence / sub-vocal segments —
|
||||
# higher temperatures produce phantom phrases that look like real transcripts.
|
||||
temperature: 0.0
|
||||
}
|
||||
)
|
||||
transcribed_text = response['text']
|
||||
|
||||
Reference in New Issue
Block a user