Whisper transcription and speech service operations

Implementation experience

• Server audio capture with Go, WebRTC and Opus, with separate tracks for each participant.
• A compressed packet journal with durable writes, segment limits and recovery after restart.
• A map between file time and conversation time to place transcription in the session.
• Transcription recovery for existing media with increasing retry intervals and an explicit waiting state.

The challenge

Speech services must meet quality, latency and concurrency requirements. When they share a GPU, memory management and model handoffs also need coordination.

Approach

• Assess models and quality using representative product audio.
• Plan integration, concurrency and memory use.
• Define monitoring, updates and recovery with the team.

Consulting scope

The proposal can include assessment, deployment, API integration and ongoing operations, with limits defined for the product's workload.

Latency and accuracy must be measured in the company's language and audio environment. Acceptance criteria are agreed before deployment.

Application in your company

Suitable for companies that need to integrate transcription and speech generation into their product, with documented operations, capacity control and recovery procedures.