ADR-006: Context-Aware Model Selection¶
- Status: Accepted
- Date: 2026-01-11
- Authors: Podcast Scraper Team
- Related RFCs: RFC-010
- Related PRDs: PRD-008
Context & Problem Statement¶
OpenAI Whisper provides generic multilingual models (base, small, etc.) and optimized English-only models (base.en, small.en). The English models are significantly more accurate and faster for English-only audio.
Decision¶
The system automatically promotes requested models to their English variants if:
- The detected or configured language is English (
en). - The user requested a model size that has an
.envariant (tiny, base, small, medium). - The user hasn't already explicitly specified a variant.
Rationale¶
- Quality by Default: Most users want the best accuracy for their language without needing to know technical model suffixes.
- Resource Efficiency: English models are typically faster and smaller than their multilingual counterparts.
Alternatives Considered¶
- Manual Selection Only: Rejected as it places too much cognitive load on the user.
- Multilingual Default: Safe, but leaves significant performance and accuracy on the table for the primary use case (English podcasts).
Consequences¶
- Positive: Better transcription quality "out of the box."
- Negative: Can be surprising to users who explicitly wanted the multilingual model for some reason (e.g., occasional non-English segments).