Which audio and rerank endpoints are supported?

POST /v1/audio/transcriptions, /v1/audio/translations, /v1/audio/speech and /v1/rerank. Transcription and speech work with OpenAI and Groq; rerank works with Cohere, Jina and Voyage. Add the provider's key in Settings and point your SDK at Cognocient.

Per what the provider bills: whisper models per second of audio, gpt-4o-transcribe models per token (read from the response), text-to-speech per character, rerank per search (Cohere) or per token (Jina, Voyage). Each response carries an x-cog-cost-basis header saying how the cost was computed.

What if a model has no price on file?

It is priced at a stated conservative rate and the cost basis is marked fallback, so the number is visibly an estimate. It is never silently recorded as zero.

Do budgets, guardrails and the Emergency Freeze apply?

Yes. The same authentication, rate limits, freeze, budget check and cost logging as chat. Text sent for speech and rerank is scanned by your PII and secret guardrails; audio files themselves are not inspected.

Which plans include them?

All of them, like embeddings and image generation. Free plan calls count toward the Free spend cap and per-minute rate limit.

Can I use audio transcription, speech and rerank through Cognocient?

Cognocient proxies /v1/audio/transcriptions, /v1/audio/translations, /v1/audio/speech and /v1/rerank with the same budgets, guardrails, freeze and cost tracking as chat. Available on every plan.

These endpoints follow the OpenAI request format, so the OpenAI SDK works unchanged with base_url="https://api.cognocient.com/v1" and your Cognocient key.

Transcription and translation

from openai import OpenAI
 
client = OpenAI(base_url="https://api.cognocient.com/v1", api_key="<your proxy key>")
 
with open("meeting.mp3", "rb") as f:
    result = client.audio.transcriptions.create(model="whisper-1", file=f)
print(result.text)

Model	Provider	Billed by
`whisper-1`	OpenAI	Second of audio
`gpt-4o-transcribe`, `gpt-4o-mini-transcribe`	OpenAI	Tokens reported in the response
`whisper-large-v3`, `whisper-large-v3-turbo`	Groq	Second of audio

Files up to 25 MB. Send response_format as usual: json, text, verbose_json, srt or vtt. To meter cost accurately Cognocient asks the provider for the richest format it supports and rebuilds the format you asked for.
For SRT and VTT the audio length is read from the last caption timestamp.
/v1/audio/translations (to English) works the same way.

Speech

audio = client.audio.speech.create(model="tts-1", voice="alloy", input="Your report is ready.")
audio.write_to_file("out.mp3")

tts-1, tts-1-hd and gpt-4o-mini-tts (OpenAI), playai-tts and canopylabs/orpheus-* (Groq). Input is limited to 5,000 characters. Cost is per character; for gpt-4o-mini-tts (which bills audio output by time) the audio length is estimated from the text and the basis is marked estimated.

Rerank

import httpx
r = httpx.post("https://api.cognocient.com/v1/rerank",
    headers={"Authorization": "Bearer <your proxy key>"},
    json={"model": "rerank-v3.5", "query": "refund policy",
          "documents": ["Returns within 30 days…", "Shipping takes 3-5 days…"], "top_n": 1, "return_documents": True})

The request and response follow Cohere's shape whichever provider serves it. Provider is chosen from the model name: rerank-v3.5, rerank-english-v3.0 and similar go to Cohere; jina-reranker-* to Jina; and Voyage's rerank-2, rerank-2.5, rerank-3 (or an explicit voyage/rerank-2) to Voyage. Up to 1,000 documents per request.

Add the provider key (Cohere, Jina or Voyage) under Settings → AI Providers first. A request for a model whose provider has no key returns a clear 400 naming the provider.

Cost and attribution

Every call is logged with your X-Cost-* attribution tags and appears in the dashboard with the endpoint that served it. The x-cog-cost-basis response header tells you how the cost was derived: seconds, tokens, characters, searches, estimated, fallback (no price on file) or unmetered (the provider returned neither a duration nor usage, so nothing could be honestly charged).

Limits

Streaming transcription and realtime audio are not supported.
Other providers (Deepgram, ElevenLabs, Azure Speech) are not supported yet.