From OpenAI / Whisper

POST /v1/audio/transcriptions is OpenAI-compatible on purpose — for batch, migration is a URL and a model prefix:

https://api.openai.com/v1/audio/transcriptions
https://api.speechrouter.ai/v1/audio/transcriptions
 -F model=whisper-1
 -F model=openai/whisper-1

response_format, language, and prompt (via provider_params) behave as you expect. Your OpenAI models are all here — whisper-1, gpt-4o-transcribe, gpt-4o-mini-transcribe, gpt-4o-transcribe-diarize — at OpenAI’s list prices.

What you gain immediately

• True streaming. OpenAI realtime needs base64-in-JSON WebSocket framing and 24 kHz audio; SpeechRouter’s /v1/listen takes binary PCM at your rate and resamples for OpenAI behind the scenes — or streams to any of the other providers with the same code.
• Word timestamps beyond whisper-1. gpt-4o-transcribe is json-only at OpenAI; on SpeechRouter, switching to a word-timing model (assemblyai/universal-2, groq/whisper-large-v3) is a slug edit.
• srt/vtt everywhere — including models where OpenAI doesn’t offer them (the gateway synthesizes cues from word timings).
• 250 MB uploads vs OpenAI’s 25 MB.
• Failover. Whisper has bad days; your transcripts shouldn’t.

Whisper at Groq prices

The same whisper-large-v3 weights run on Groq’s hardware at ~200× realtime for $0.111/hr ($0.04/hr turbo) — model=groq/whisper-large-v3. One slug, an order of magnitude cheaper for batch.