bug: OpenAI gpt-4o-transcribe live transcription issues

[devin-ai-integration]

Description

A user reported issues when using OpenAI gpt-4o-transcribe for live transcription.

Analysis

The default live model for OpenAI is gpt-4o-transcribe (crates/owhisper-client/src/providers.rs:275).

Missing word-level timestamps in live mode

In the live adapter (crates/owhisper-client/src/adapter/openai/live.rs), build_transcript_response creates words by splitting whitespace but sets start: 0.0 and end: 0.0 for every word — OpenAI's Realtime API does not provide word-level timestamps for transcription. This means every response has start: 0.0, duration: 0.0, which could cause issues with:

• Subtitle display and word highlighting
• Transcript alignment
• Diarization / speaker assignment

Other potential issues

• Server-side VAD settings (threshold=0.5, silence_duration=500ms) may be too aggressive or not aggressive enough for certain environments
• The channel_index is hardcoded to [0, 1] regardless of actual channel mode
• The include field requests item.input_audio_transcription.logprobs which may not be supported for all models
• For batch transcription, gpt-4o-transcribe correctly uses json format instead of verbose_json, but this means no word-level timing data

Relevant Files

• crates/owhisper-client/src/adapter/openai/live.rs — OpenAI Realtime API live adapter
• crates/owhisper-client/src/adapter/openai/batch.rs — OpenAI batch transcription adapter
• crates/owhisper-client/src/providers.rs — Provider configuration and default models
• crates/owhisper-client/src/adapter/parsing.rs — Word builder and time span calculation