---
title: Qwen, MiniMax and Mistral speech
description: Route speech to Qwen TTS on Model Studio and MiniMax T2A, and transcription to Mistral's Voxtral, through the orchestrator's audio API.
section: voice
order: 30
audience: [operator, builder]
sources: [services/go-orchestrator/internal/voice/dashscope.go, services/go-orchestrator/internal/voice/minimax.go, services/go-orchestrator/internal/voice/native.go, config/orchestrator.yaml, config/prices.yaml]
features: []
---

Three cloud providers reach the orchestrator's audio API through their own wire formats. Clients still send
OpenAI's requests to `POST /v1/audio/speech`, `GET /v1/audio/voices` and `POST /v1/audio/transcriptions`.

| Provider | Upstream kind | Models | What it serves |
|---|---|---|---|
| Qwen TTS (Alibaba Cloud Model Studio) | `dashscope` (the `qwen` upstream) | `qwen3-tts-flash` | speech, voices |
| MiniMax | `minimax` | `speech-2.8-hd`, `speech-2.8-turbo` (and the legacy 2.6, 02 and 01 families) | speech, voices |
| Mistral | `mistral` | `voxtral-mini-latest` | transcription |

## Qwen TTS

The shipped `qwen` upstream lists `qwen3-tts-flash`, and the `china` profile's `venom-speak` alias routes to it.
The orchestrator always uses Model Studio's streaming mode, so audio arrives as it is made:

- `response_format: "wav"` returns a 24 kHz mono 16-bit WAV.
- `response_format: "pcm"` returns the same samples without the header, labeled
  `audio/pcm;rate=24000;channels=1`.
- Any other format is a `400 unsupported_response_format`.

`voice` defaults to `Cherry`; `GET /v1/audio/voices?model=qwen/qwen3-tts-flash` lists the 45 system voices with
their languages. `venom_options.language` (`en`, `zh`, `ja`, ...) sets the language. `speed` and `instructions`
are not sent. Price: $0.10 per 10,000 characters (international).

## MiniMax

```yaml
upstreams:
  - name: minimax
    kind: minimax
    base_url: https://api.minimax.io/v1
    api_key_env: MINIMAX_API_KEY
    models: [MiniMax-M3, speech-2.8-turbo, speech-2.8-hd]   # chat and speech on one upstream
```

`voice` is required: take a `voice_id` from `GET /v1/audio/voices?model=minimax/speech-2.8-turbo`. Formats are
`mp3`, `pcm`, `flac`, `wav` and `opus` (`aac` is a 400). `speed` is clamped to MiniMax's 0.5 to 2.
`venom_options.language` sets `language_boost`, and `venom_options.sample_rate` takes 8000 to 44100 (pcm and wav
default to 24000). With `stream: true` (or `stream_format: "sse"`) audio comes from MiniMax's event stream;
without it, from one answer. MiniMax's error codes become HTTP statuses: rate limits are 429, a bad key is 401,
an empty balance is 402, and bad input is 400. Price: $100 (hd) or $60 (turbo) per 1M characters.

`venom_options.provider` reaches anything else: `{"voice_setting": {"emotion": "calm"}}` is merged into the
request MiniMax gets, keeping the voice.

## Mistral Voxtral

Voxtral takes OpenAI's multipart form, so an upstream of kind `mistral` passes transcriptions through unchanged:

```yaml
upstreams:
  - name: mistral
    kind: mistral
    base_url: https://api.mistral.ai/v1
    api_key_env: MISTRAL_API_KEY
    models: [voxtral-mini-latest]
```

The billed audio (`usage.prompt_audio_seconds`) appears in the request log as `audio_ms`. Voxtral has no bundled
price yet; set one with `pricing.models: {voxtral-mini-latest: {per_second: ...}}`.
