DEVELOPER GUIDE

From audio to API.
Start here.

Speechinfra exposes model-specific endpoints for speech and audio processing. Public documentation does not require an account. Executing a model does.

1. Create your account

Register, verify your email and sign in. A default project API key is created for your verified account. Copy it from Keys in Console and store it as SPEECH_API_KEY in your server environment. Do not commit it to source control. Running a model requires sufficient balance; contact us if you need funding assistance.

2. Choose a model

Browse the public catalog for current prices, languages, request parameters and model-specific examples. File endpoints accept clips up to 30 seconds and 20 MiB. Alignment also requires a matching transcript. Enhancement returns audio, not a transcript. Realtime uses WebSocket rather than a file POST.

3. Send a short recording

curl --fail-with-body https://api.speechinfra.com/v1/inference/whisper-large-v3-turbo-1 \
  -H "Authorization: Bearer $SPEECH_API_KEY" \
  -F "file=@clip.wav"

Use a clip you have permission to process. The model page contains Python and JavaScript examples and its response schema. These examples are documentation, not requests executed by this page.

Billing and errors

Cost is based on audio duration and the selected model’s current rate, not file size. File uploads reach the server before the duration is measured and the cost reserved. The model is not started if the reservation cannot be funded. Failed file processing releases its reservation. Realtime reserves funds before sending each additional chunk and charges for audio already sent, even if the session later disconnects.

401: missing or invalid authentication.
402: insufficient balance; fund the project before retrying.
413 / 422: check file limits and request parameters.
429: request or concurrency limit; retry with backoff.
5xx: temporary processing or service error. Avoid blindly retrying requests whose outcome is unknown.

Record the response’s x-request-id for support. Prices in the catalog and Console share the same database. A model without a configured price uses the $1/audio-minute fallback; check its displayed rate before execution.

Frequently asked questions

What is Speechinfra?

Speechinfra provides API access to speech and audio models for transcription, realtime speech recognition, speaker diarization, speech alignment, enhancement and voice activity detection. Choose a specific model for each request.

How do I start?

Create an account and verify your email. Copy your project API key from Console for server-side API calls. The web playground uses your signed-in session; you do not paste a key into it. A funded balance is required to run a model.

Which files can I send?

The direct file API accepts short audio or video clips, up to 30 seconds and 20 MiB. WAV is used in our examples. Accepted containers and additional settings depend on the selected model. Realtime uses binary mono PCM16 at 16 kHz, not uploaded compressed files.

Which languages and speaker features are supported?

Support varies by model. Each public model page lists the language codes, automatic language detection support and request parameters exposed by our API. Do not assume every model supports every language, speaker labels or word timestamps.

How is usage billed?

Prices are in USD per minute of audio and vary by model. For file processing, the server measures duration and reserves the cost before starting the model. Insufficient funds return HTTP 402. Failed file processing releases the reservation. Realtime reserves incrementally and charges for audio already sent.

Can I use the API without opening Console each time?

Yes. Send your project key in the Authorization: Bearer header from your application backend. Never expose provider credentials or embed a project key in a public frontend bundle.

Need help?

Contact our team with your model ID and request ID. Do not send passwords, API keys or sensitive recordings.