Voice AI Transcription Reliability

Freelance Brief: Voice Transcription Reliability We’re looking for a freelance engineer to investigate and improve a production voice AI system where occasional user utterances—particularly short or overlapping speech—may not be transcribed. Initial investigation suggests the issue may involve interactions between voice activity detection (VAD), interruption handling, streaming speech-to-text connections, and audio buffering during recovery. Scope - Reproduce and isolate the missing-utterance scenarios. - Analyze VAD, interruption, and streaming STT behavior. - Identify the root cause across application, SDK, and provider boundaries. - Improve diagnostics without logging sensitive conversation content. - Propose and implement a reliable fix with automated tests. - Validate the solution under realistic latency, overlap, and failure conditions. - Document findings, trade-offs, and operational recommendations. Required skills - Strong Python and asynchronous programming experience. - Experience with real-time audio or voice AI systems. - Familiarity with streaming speech-to-text and VAD. - Understanding of buffering, retries, concurrency, and network failure recovery. - Production observability experience with logs, metrics, and distributed tracing. - Ability to debug third-party SDK and provider integrations. - Strong testing and root-cause analysis practices. Experience with LiveKit Agents, WebRTC, gRPC, and cloud speech providers is preferred.