Two production-grade fixes to the existing audio transcription service.
**Independent of the WhatsApp Calling work** — these affect every audio
attachment that goes through Whisper (voice notes, call recordings,
voicemails, etc.).
## Closes
- [PLA-151 — PR-5: Recording Upload + Transcription
Pipeline](https://linear.app/chatwoot/issue/PLA-151/pr-5-recording-upload-transcription-pipeline)
## Why this is needed
### 1. Whisper rejects payloads larger than 25 MB
OpenAI's [Whisper
API](https://platform.openai.com/docs/guides/speech-to-text) hard-caps
file uploads at 25 MB. Long audio recordings — voice notes from chatty
contacts, ~70+ min Opus call recordings — currently hit OpenAI with the
full payload and 413 (\`Payload Too Large\`). The job retries via the
existing \`Faraday::BadRequestError\` discard path, but the agent still
sees a transcription failure for an attachment we knew was too big up
front.
This PR adds a pre-flight \`audio_too_large?\` check via the blob's
\`byte_size\` and returns a controlled error without hitting OpenAI. The
audio attachment is preserved (agents can still listen), only the
transcription is skipped.
### 2. Whisper hallucinates on silence at non-zero temperature
At \`temperature: 0.4\` (the previous value), Whisper produces
well-documented hallucinated repeats on silence and near-silent segments
— e.g. \`Oh, dear. Oh, dear. Oh, dear.\` filling the transcript. This
shows up in real recordings whenever there's a hold or quiet moment.
\`temperature: 0.0\` matches OpenAI's recommended default for
transcription and eliminates the spirals.
Reference:
[openai/whisper#928](https://github.com/openai/whisper/discussions/928),
[openai-python#1010](https://github.com/openai/openai-python/issues/1010).
## Are WhatsApp call recordings already handled?
Yes — by the existing pipeline, **before this PR**:
\`\`\`
Browser MediaRecorder → upload_recording (PR-4)
→ @call.message.attachments.create!(file_type: :audio, ...)
→ Enterprise::Concerns::Attachment#enqueue_audio_transcription
(after_create_commit hook)
→ Messages::AudioTranscriptionJob.perform_later(attachment.id)
→ Messages::AudioTranscriptionService → Whisper
\`\`\`
The \`after_create_commit\` hook already fires for every audio
attachment regardless of source. PR-4's \`upload_recording\` endpoint
creates the attachment; the existing job/service take it from there. No
new wiring needed.
This PR just makes the existing service more robust:
- Calls longer than ~70 min (Opus 48 kbps) no longer 413 against OpenAI
- Quiet recordings no longer produce hallucinated transcripts
## How to test
\`\`\`ruby
# In rails console with a real audio attachment:
service = Messages::AudioTranscriptionService.new(Attachment.audio.last)
# Normal-sized audio: unchanged behaviour
service.perform # => { success: true, transcriptions: ... }
# Large audio: new guard returns error instead of 413-ing OpenAI
allow(attachment.file.blob).to
receive(:byte_size).and_return(30.megabytes)
service.perform # => { error: 'Audio too large for Whisper' }
\`\`\`
Existing transcription specs cover the happy path; one new spec
exercises the byte-limit guard.
## Risk
Low. Both changes are pre-flight guards or parameter values — they
reduce the surface of OpenAI calls that can fail. Failure to transcribe
is already non-fatal (the audio attachment is preserved either way).
