OpenAI‑compatible speech & embeddings on Tenstorrent
OpenAI-compatible speech-to-text (Whisper large-v3) and text embeddings (BGE-M3), billed per use from your QuietForge wallet. Create a key on your account page.
from openai import OpenAI client = OpenAI(base_url="https://quietforge-api.badgerbite.io/v1", api_key="qfk_...") # Whisper large-v3 speech-to-text with open("meeting.mp3", "rb") as f: text = client.audio.transcriptions.create( model="whisper-1", file=f) # BGE-M3, 1024-dimensional vectors vecs = client.embeddings.create( model="bge-m3", input=["hello", "world"])
Whisper large-v3
Transcribe or translate audio into English. JSON, text, SRT, VTT or verbose JSON.
BGE-M3
Normalised 1024-dimensional vectors, up to 256 inputs per request.
Prepaid, per use
We reserve an estimate, then charge what was actually used. No subscription.
Endpoints
https://quietforge-api.badgerbite.io/v1/audio/transcriptionsMultipart file, model
(whisper-large-v3 or whisper-1), optional language, prompt,
response_format (json, text, srt, vtt, verbose_json)
https://quietforge-api.badgerbite.io/v1/audio/translationsThe same, translating into English
On both audio endpoints, language is the source language of the audio (an ISO-639-1 code such
as fr). Without it the audio is treated as English, because automatic language detection is not
available yet: send language for any non-English audio, including audio you want translated.
https://quietforge-api.badgerbite.io/v1/embeddingsJSON {"model": "bge-m3", "input": ["..."]}, 1024-dimensional normalised vectors
https://quietforge-api.badgerbite.io/v1/modelshttps://quietforge-api.badgerbite.io/v1/jobs/{id}Prices
| Model | Job | Price |
|---|---|---|
bge-m3 | embedding | $0.00002 per started 1k tokens |
whisper-large-v3 | transcription | $0.003 per started audio minute |
whisper-large-v3 | translation | $0.003 per started audio minute |
We reserve an estimate when a request starts and charge what was actually used. Failed requests are free.
Examples
curl https://quietforge-api.badgerbite.io/v1/audio/transcriptions -H "Authorization: Bearer $QF_KEY" \ -F file=@meeting.mp3 -F model=whisper-1 -F response_format=srt
from openai import OpenAI client = OpenAI(base_url="https://quietforge-api.badgerbite.io/v1", api_key="qfk_...") client.embeddings.create(model="bge-m3", input=["hello", "world"])
Long audio, async and webhooks
Requests wait up to 60 s. A synchronous call that runs out of time does not
fail: it returns 202 with a job object ({"id", "object", "status", "poll"}) instead of the
transcript, and the job keeps running. The OpenAI SDKs don't expect that response, so for long audio send
sync=false (with the SDK, extra_body={"sync": "false"}) and poll
/v1/jobs/{id}, or add a callback_url. We POST the job to your URL with the header
X-QF-Signature: t=…,v1=…, an HMAC-SHA256 of "{t}.{body}" using your key's webhook
secret. Check the signature and reject stale timestamps:
import hmac, hashlib, time t, v1 = (p.split("=", 1)[1] for p in header.split(",")) ok = hmac.compare_digest(v1, hmac.new(secret.encode(), f"{t}.".encode() + body, hashlib.sha256).hexdigest()) if not ok or abs(time.time() - int(t)) > 300: reject()
Errors about a job that already exists carry X-QF-Job-Id and x-should-retry: false, so
the SDK doesn't retry and start a second job: check /v1/jobs/{id} first.
Webhooks are delivered at least once. If a delivery times out or fails, we retry it, so treat a repeated payload
with the same job id and state as a duplicate.
Limits
Audio up to 25 MiB and 2 hours; a prompt of up to 224 tokens. Up to 256 embedding inputs per
request, each at most 8192 tokens (longer inputs are rejected, not truncated).