SenseAudio-ASR

Build and troubleshoot SenseAudio speech recognition integrations, including HTTP transcription (/v1/audio/transcriptions), realtime WebSocket ASR (/ws/v1/audio/transcriptions), audio quality analysis (/v1/audio/analysis), and recognition record queries (/v1/audio/records). Use this whenever user asks for speech-to-text, diarization, translation, streaming ASR, or ASR model/parameter selection.

Use this skill for all SenseAudio speech recognition tasks.

Credential source: read the API key from SENSEAUDIO_API_KEY and send it only in the Authorization: Bearer ... header. Do not place API keys in query parameters, logs, transcripts, or saved examples.

Read First

• references/asr.md

Workflow

1. Pick recognition mode:

• HTTP file transcription for offline audio.
• WebSocket for realtime streaming microphone/audio chunks.
• Audio analysis for noise and quality checks before recognition.
• Records query for recent recognition history lookup.

2. Choose model by feature needs:

• Lite for low-cost basic transcription.
• ASR for streaming, translation, diarization, sentiment, and timestamps.
• Pro when diarization plus explicit max_speakers control is needed.
• DeepThink for streaming, translation, and intelligent editing; do not send language, diarization, sentiment, timestamps, ITN, or punctuation controls.

3. Build minimal request:

• Required auth, file/audio format, model.
• Add optional controls only when needed.
• Keep uploaded files at or below 10MB; split longer audio before sending.

4. Validate compatibility:

• Check model-parameter support before sending.
• Enforce WS pcm / 16000Hz / mono requirements.
• For HTTP stream=true, expect SSE text deltas only, not structured verbose fields.

5. Parse robustly:

• Handle JSON/text/verbose/SSE forms.
• Handle WS terminal events and failures.
• Treat returned audio URLs, api_key, session_id, and trace_id as sensitive operational data.