Transcribe multichannel speech-to-text

Turn on multichannel when the audio already has discrete channels (call center agent on ch0, customer on ch1). Batch returns a channels array; streaming tags partials with channel_index and emits transcript.done once per channel. Base flows: Transcribe audio with the Voice API and Stream speech to text over WebSocket.

Batch REST

POST https://api.x.ai/v1/stt. multichannel is a boolean (default false). When true, the response includes channels: each item has index, text, and words.

For raw PCM/mulaw/alaw, pass channels as an integer from 2 to 8. Container formats (WAV, MP3, FLAC, OGG, WebM, MP4, M4A) auto-detect channel count.

curl -X POST https://api.x.ai/v1/stt \
  -H "Authorization: Bearer $XAI_API_KEY" \
  -F multichannel=true \
  -F file=@/path/to/stereo-call.wav

Max batch file size is 500 MB. Up to 8 channels.

Streaming WebSocket

multichannel=true requires channels ≥ 2. Multichannel is unsupported with encoding=opus. Send interleaved PCM (L,R,L,R…).

wss://api.x.ai/v1/stt?sample_rate=16000&encoding=pcm&multichannel=true&channels=2&interim_results=true

• transcript.partial includes channel_index.
• transcript.done fires once per channel.
• Finalize all channels with {"type":"Finalize"}. Finalize one channel with {"type":"Finalize","channel":N}.

Chunk size note: stereo PCM16 at 16 kHz for 100 ms is 6400 bytes (sample_rate * channels * bytes_per_sample * duration_s → 16000 * 2 * 2 * 0.1).

Proxy the WebSocket through your backend so the API key stays off the client.

Speaker labels on mixed mono

For a single mixed track where you need speaker integers on each word, use Enable speaker diarization on speech-to-text.

Pitfalls

• multichannel=true with encoding=opus is rejected on streaming.
• Raw audio needs an explicit channels value (2–8); containers auto-detect.
• Cap is 8 channels. Batch max file size is 500 MB.
• Console API credits are separate from SuperGrok's weekly pool on grok.com.