Audio formats

Streaming input

Declare what you’ll send in the connection params; the gateway rejects model/encoding combinations that can’t work at connect time, not mid-stream:

Param	Default	Notes
encoding	linear16	Raw 16-bit little-endian PCM — supported everywhere
sample_rate	16000	16 kHz is the sweet spot for every provider
channels	1	Mono unless the model supports more

linear16 @ 16 kHz mono is the universal safe choice and what the SDK examples assume. Telephony audio (mulaw @ 8 kHz) and compressed input (opus, ogg-opus) are supported where the provider supports them — each provider page lists its encodings.

Chunking

Send binary frames of 20–250 ms of audio. Two provider quirks the gateway absorbs so you don’t have to:

• Minimum chunk sizes (e.g. AssemblyAI rejects < 50 ms): sub-minimum chunks are coalesced before forwarding.
• Real-time pacing (some providers close the socket if audio arrives faster than real time): the gateway re-paces bursts.

So: send audio as it arrives from your source, at whatever cadence — the gateway does the per-vendor conditioning.

Sample-rate conversion

Providers with fixed input rates (e.g. OpenAI realtime wants 24 kHz) get gateway-side resampling from your declared rate. Declare what you actually have; never resample client-side.

Batch input

POST /v1/audio/transcriptions accepts whatever container the model accepts — WAV, MP3, FLAC, OGG, M4A and friends — up to 250 MB. The gateway forwards your file’s content type; no transcoding is applied.