ADR 002: Use OpenAI Whisper for Transcription

Status

Accepted

Context

Audio transcription is critical for the pipeline. Requirements:

• High accuracy (>95% for clear audio)
• Timestamp alignment
• Multi-language support
• Cost efficiency

Decision

Use OpenAI Whisper API (whisper-large-v3) for transcription.

Consequences

Positive

• State-of-the-art accuracy
• Supports 99 languages
• Returns word-level timestamps
• Cost-effective ($0.006/minute)
• Handles various audio qualities
• Speaker diarization available (with extensions)

Negative

• Requires internet connection
• API rate limits (50 RPM)
• Audio file size limits (25MB)
• Potential latency for long files

Alternatives Considered

Alternative	Pros	Cons
Google Speech-to-Text	Enterprise features	Higher cost
AWS Transcribe	AWS integration	Accuracy lower
Local Whisper	No API dependency	Requires GPU, slower
AssemblyAI	Good accuracy	Higher cost
Deepgram	Fast, accurate	Higher cost

Implementation

import openai

with open(audio_path, "rb") as audio_file:
    transcript = openai.audio.transcriptions.create(
        model="whisper-large-v3",
        file=audio_file,
        response_format="verbose_json",
        timestamp_granularities=["word", "segment"]
    )

Notes

For files >25MB, split using FFmpeg before sending to API. Implement retry logic with exponential backoff for rate limits.