---
title: ElevenLabs, Fish Audio and BytePlus Seed Speech
description: Route text to speech and transcription to ElevenLabs, Fish Audio or BytePlus Seed Speech through the orchestrator's OpenAI-shaped audio API.
section: voice
order: 30
audience: [user, operator, builder]
sources: [services/go-orchestrator/internal/voice/elevenlabs.go, services/go-orchestrator/internal/voice/fishaudio.go, services/go-orchestrator/internal/voice/byteplus.go, services/go-orchestrator/internal/voice/sttformat.go, docs/configuration.md]
features: []
---

Three cloud speech providers have upstream kinds of their own: `elevenlabs`, `fishaudio` and `byteplus-speech`.
Clients keep using the OpenAI audio endpoints, `POST /v1/audio/speech`, `GET /v1/audio/voices` and
`POST /v1/audio/transcriptions`, and the orchestrator translates each request to the provider's API. Aliases,
fallback, breakers, key budgets and the request log work as they do for other models. The text and audio you
send to these providers leave your machine.

## Configure the upstreams

These providers don't list models, so name the ones you use:

```yaml
upstreams:
  - name: elevenlabs
    kind: elevenlabs
    base_url: https://api.elevenlabs.io/v1
    api_key_env: ELEVENLABS_API_KEY
    models: [eleven_flash_v2_5, eleven_multilingual_v2, scribe_v2]
  - name: fish
    kind: fishaudio
    base_url: https://api.fish.audio
    api_key_env: FISH_API_KEY
    models: [s2.1-pro, transcribe-1]
  - name: byteplus
    kind: byteplus-speech
    base_url: https://voice.ap-southeast-1.bytepluses.com
    api_key_env: BYTEPLUS_SPEECH_API_KEY
    models: [seed-tts-2.0, seed-asr-fast]
routes:
  - alias: venom-speak
    targets:
      - {upstream: elevenlabs, model: eleven_flash_v2_5}
pricing:
  models:
    elevenlabs/eleven_flash_v2_5: {per_1m_chars: 50}
```

None of the three has a list price in the bundled price table, because they bill through plans and credits. Set
`per_1m_chars` (USD per 1M input characters) for speech and `per_second` (USD per second of audio) for
transcription so that spend views and key budgets count them.

## What each provider supports

| Kind | Auth | Speech formats | Voice | Transcription |
|---|---|---|---|---|
| `elevenlabs` | `xi-api-key` header | mp3, opus, wav, pcm (8–48 kHz) | required, a voice id | Scribe (`scribe_v2`) |
| `fishaudio` | bearer token | mp3, opus, wav, pcm (8–44.1 kHz) | optional, a voice model id | `transcribe-1`, `transcribe-1-pro` |
| `byteplus-speech` | `X-Api-Key` header (a Seed Speech key, not a ModelArk key) | mp3, opus, pcm | required, a speaker id | the flash recognizer (`seed-asr-fast`), wav, mp3 or ogg files |

If you ask for a format the provider can't produce, or leave out a voice that the provider requires, the
orchestrator answers `400` with the code `unsupported_by_provider` and never calls the provider. A pcm answer gives
its sample rate in `Content-Type: audio/pcm;rate=…` and in `x-venom-sample-rate`. Ask for another rate with
`venom_options.sample_rate`.

The OpenAI fields are mapped to each provider's own options:

- `speed` becomes ElevenLabs' `voice_settings.speed`, Fish Audio's `prosody.speed`, or BytePlus's `speech_rate`.
- `venom_options.language` becomes `language_code` (ElevenLabs), `language` (Fish Audio) or `explicit_language` (BytePlus).
- `venom_options.provider` is merged last into the provider's request body, so any other option is reachable.
- With `stream: true`, ElevenLabs uses its streaming endpoint. BytePlus streams base64 chunks, which the
  orchestrator decodes to raw audio as they arrive.

## Voices

`GET /v1/audio/voices?model=<alias>` lists:

- ElevenLabs: the voices in your library.
- Fish Audio: your own voice models, or the most used public ones if you have none.
- BytePlus: a fixed subset of the official Seed Speech TTS 2.0 voices, because BytePlus has no voice-list API.

Each voice carries `provider` (the upstream name) and `model`.

## Transcription

A transcription through any of the three answers `response_format` `json`, `text`, `verbose_json`, `srt` or `vtt`.
The orchestrator builds each format from the provider's timed transcript and logs the billed audio as `audio_ms`.
None of the three serves `/v1/audio/translations`. BytePlus's streaming recognition and bidirectional speech run
over WebSocket and aren't offered here.
