From AssemblyAI

Your AssemblyAI models are on SpeechRouter (assemblyai/universal-3-5-pro, universal-streaming-*, universal-2) at AssemblyAI’s own list prices — so migration is a wire-format change, not a model change. You gain the other eleven providers, mid-stream failover, and one bill.

Streaming

AssemblyAI v3	SpeechRouter
wss://streaming.assemblyai.com/v3/ws?...	wss://api.speechrouter.ai/v1/listen?model=assemblyai/universal-3-5-pro
Authorization: <key> header	Bearer header or ['bearer', key] subprotocol
Turn messages, accumulate by turn_order	transcript events with plain is_final
end_of_turn && turn_is_formatted dance	just is_final: true (the gateway dedupes the formatted/unformatted double)
chunk sizing 50–1000 ms + realtime pacing rules	send at any cadence — the gateway coalesces and re-paces
letter speakers "A"	integer speakers 0
temp tokens via /v3/token	POST /v1/tokens

Vendor knobs (end_of_turn_confidence_threshold, format_turns, PII redaction …) keep working via typed provider_params.

Batch

POST /v2/transcript + polling becomes one synchronous call:

curl -s https://api.speechrouter.ai/v1/audio/transcriptions \
  -H "Authorization: Bearer $SPEECHROUTER_API_KEY" \
  -F model=assemblyai/universal-2 -F diarization=true \
  -F response_format=verbose_json -F file=@call.mp3

No upload_url step, no polling loop — the response is the transcript.

Keep your AssemblyAI account (optional)

BYOK: paste your AssemblyAI key in the console and sessions on their models run on your account, billed by SpeechRouter at $0.00.