---
title: Voice API on Venom Cloud
description: Text to speech, voice lists and transcription on Venom Cloud through your own provider keys, with the orchestrator's OpenAI-shaped audio API.
section: cloud
order: 215
audience: [user, builder, operator]
sources: [services/nodejs-cloud/app/v1/audio, services/nodejs-cloud/lib/voice/serve.ts, services/nodejs-cloud/lib/voice/adapters.ts, services/nodejs-cloud/lib/voice/stt.ts, services/nodejs-cloud/lib/providers/registry.ts, services/nodejs-cloud/migrations/0114_voice_usage.sql, config/prices.yaml, docs/cloud-mesh.md]
features: [cloud.v1-audio]
---

Venom Cloud serves the same audio API as the orchestrator ([Speech API](/docs/voice/speech-api.md)), with your
Cloud API key and the provider keys on the [Providers page](/docs/cloud/console-providers.md#voice):

| Endpoint | What it does |
|---|---|
| `POST /v1/audio/speech` | Text to speech. The audio streams back as the provider sends it. |
| `GET /v1/audio/voices?model=` | The voices of every text-to-speech target of the model, each tagged with `provider` and `model`. |
| `POST /v1/audio/transcriptions` | Speech to text from a multipart upload (`file`, `model`, optional `language`, `response_format`). |
| `POST /v1/audio/translations` | Speech to English text, on providers that offer it (OpenAI-compatible and Mistral). |

```bash
curl https://cloud.usevenom.dev/v1/audio/speech \
  -H "Authorization: Bearer $VENOM_API_KEY" -H 'content-type: application/json' \
  -d '{"model": "eleven_flash_v2_5", "input": "Hello from Venom.", "voice": "<voice id>"}' \
  -o hello.mp3
```

## Providers {#providers}

| Provider | Speech | Voices | Transcription | Notes |
|---|---|---|---|---|
| OpenAI (`openai`) | `gpt-4o-mini-tts`, `tts-1`, `tts-1-hd` | built-in list | `whisper-1`, `gpt-4o-transcribe`, `gpt-4o-mini-transcribe` | translations too |
| ElevenLabs (`elevenlabs`) | `eleven_v3`, `eleven_multilingual_v2`, `eleven_flash_v2_5`… | your voice library | `scribe_v2` | `stream: true` uses ElevenLabs' stream endpoint |
| Fish Audio (`fishaudio`) | `s2.1-pro`, `s1` | your voices, else popular ones | `transcribe-1` | |
| BytePlus Seed Speech (`byteplus-speech`) | `seed-tts-2.0`, `seed-tts-1.0`, `seed-icl-2.0` | a few official voices | `seed-asr-fast` (wav, mp3, ogg) | mp3, opus or pcm |
| MiniMax (`minimax`) | `speech-2.8-hd`, `speech-2.8-turbo` | system voices | | |
| Model Studio (`qwen`) | `qwen3-tts-flash` | built-in list | | wav or pcm, 24 kHz |
| Mistral (`mistral`) | | | `voxtral-mini-latest` | translations too |
| Your own server (`custom`) | OpenAI-compatible (Kokoro-FastAPI, Speaches, LocalAI) | its `/audio/voices` | its `/audio/transcriptions` | |

On Venom Cloud itself, the platform's own OpenAI and Model Studio keys also serve the OpenAI models above and
`qwen3-tts-flash` when your organization uses Venom's keys. CosyVoice and Model Studio's recognizers are
WebSocket APIs and are not offered here.

## Requests and answers {#requests}

Speech takes OpenAI's body: `model`, `input`, `voice` (an id, or `{"id": …}`), `response_format` (`mp3` by
default, `opus`, `aac`, `flac`, `wav`, `pcm`), `speed` (0.25 to 4), `instructions`, `stream`, `stream_format`
(`audio` or `sse`) and Venom's `venom_options`: `language`, `sample_rate`, and `provider` (merged into the
provider's own request as is). Each provider takes only some formats and rates; one it can't serve is a `400`
saying what it takes.

- `pcm` and `wav` answers carry `x-venom-sample-rate`; `pcm` is `audio/pcm;rate=<rate>;channels=1`.
- `stream_format: "sse"` answers `speech.audio.delta` events (base64 audio) and a final `speech.audio.done` with
  `usage.input_characters`.
- Every answer names the provider and model that served it: `x-venom-upstream`, `x-venom-model`.
- Refusals before any provider is called: `missing_model`, `missing_input`, `invalid_response_format`,
  `invalid_speed`, `invalid_stream_format`, `invalid_voice`, `model_not_found`, and `model_not_tts` or
  `model_not_stt` for a model that does something else.

Transcriptions answer `json` (the default), `text`, `verbose_json`, `srt` or `vtt`, whatever the provider's own
format. The upload is at most 4.5 MB on Venom Cloud (the hosting platform's request limit); send longer audio to
a self-hosted orchestrator.

## Fallback {#fallback}

A model's targets are your keys in serving order, then Venom's keys. A target that doesn't answer, answers 429
or fails with a 5xx before sending any audio is skipped for the next. A provider's refusal (for example an
unknown voice) is returned as the provider's own JSON. Once audio has started, the stream is never switched to
another provider. Voice ids belong to a provider, so a model whose targets span providers needs a voice each
provider knows.

## Daemons {#daemons}

Voice requests are served by provider keys only, never relayed to your daemons: a relayed job carries a JSON
body and daemons don't accept audio jobs yet. A daemon's own clients can still use its audio API directly.

## Usage and cost {#usage}

Speech is billed per input character and transcription per second of audio, at the list prices in
`config/prices.yaml` (OpenAI's TTS and Whisper, Qwen TTS, MiniMax speech). A model without a voice price is
logged at $0. A key with a hard budget is served only by priced targets. Each request appears in the
[log explorer](/docs/cloud/console-log-query.md) as `kind:speech` or `kind:transcription`, with the characters or seconds
it was billed for. Your organization's custom prices on the Pricing page don't apply to voice yet.
