MiniMax Speech 2.8

MiniMax Speech 2.8 is available through the OpenAI-compatible text-to-speech endpoint. The current integration supports synchronous speech generation for short narration, dubbing, and voice output.

Endpoint

POST /v1/audio/speech

Supported Models

Model	Description
minimax-speech-2.8-turbo	Low-latency speech synthesis model
minimax-speech-2.8-hd	Higher-quality speech synthesis model

Request Parameters

Parameter	Type	Required	Description
model	string	Yes	minimax-speech-2.8-turbo or minimax-speech-2.8-hd
input	string	Yes	Text to synthesize
voice	string	Yes	MiniMax voice ID, such as a system voice, cloned voice, or generated voice ID
response_format	string	No	Audio format. Default is mp3; common values include mp3, wav, flac, and pcm
speed	number	No	Speech speed, commonly 0.5 to 2; default is 1
metadata	object	No	MiniMax extension options

Basic Example

curl -X POST "https://cubicspaces.cloud/v1/audio/speech" \
  -H "Authorization: Bearer $NEW_API_KEY" \
  -H "Content-Type: application/json" \
  --output minimax-speech.mp3 \
  -d '{
    "model": "minimax-speech-2.8-turbo",
    "voice": "male-qn-qingse",
    "input": "Hello, this is a MiniMax Speech 2.8 synchronous speech generation test.",
    "response_format": "mp3",
    "speed": 1
  }'

On success, the response body is binary audio. Use --output or an equivalent client option to save the file.

Create And Use A Cloned Voice

MiniMax Speech can synthesize audio with a cloned voice. First create the cloned voice through the voice cloning endpoint and obtain voice_id. Then pass that voice_id as the voice value when calling /v1/audio/speech.

Typical workflow:

1. Prepare source audio for cloning. A clear single-speaker mp3, m4a, or wav file with low background noise is recommended.
2. Call /v1/audio/voice-cloning with the source audio URL, and optionally provide preview text and a preview model.
3. Read voice_id from the response.
4. Use that voice_id with /v1/audio/speech.

Cloning audio and retention notes:

Item	Description
Source audio	Use a publicly accessible audio URL. Supported formats include mp3, m4a, and wav
Duration	At least 10 seconds and no more than 5 minutes
File size	No more than 20 MB
Prompt audio	Optional. If clone_prompt is provided, the prompt audio should be shorter than 8 seconds, with matching text ending in punctuation
Result	The cloning workflow returns voice_id. If preview text and a preview model are provided, it usually also returns a preview audio URL
Voice retention	Fast-cloned voices may be temporary. After creating one, call that voice_id through /v1/audio/speech as soon as possible

Create a cloned voice:

curl -X POST "https://cubicspaces.cloud/v1/audio/voice-cloning" \
  -H "Authorization: Bearer $NEW_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "minimax-voice-cloning",
    "audio_url": "https://example.com/voice.mp3",
    "text": "This text is used to preview the cloned voice.",
    "preview_model": "speech-2.8-hd",
    "accuracy": 0.7,
    "need_noise_reduction": true,
    "need_volume_normalization": true
  }'

Response example:

{
  "demo_audio_url": "https://example.com/demo.mp3",
  "voice_id": "voice_abc123"
}

To improve voice similarity and stability, add clone_prompt:

{
  "model": "minimax-voice-cloning",
  "audio_url": "https://example.com/voice.mp3",
  "clone_prompt": {
    "prompt_audio_url": "https://example.com/prompt.mp3",
    "prompt_text": "This is the text that matches the prompt audio."
  },
  "text": "This text is used to preview the cloned voice.",
  "preview_model": "speech-2.8-hd"
}

Cloned voice synthesis example:

curl -X POST "https://cubicspaces.cloud/v1/audio/speech" \
  -H "Authorization: Bearer $NEW_API_KEY" \
  -H "Content-Type: application/json" \
  --output cloned-voice.mp3 \
  -d '{
    "model": "minimax-speech-2.8-hd",
    "voice": "voice_abc123",
    "input": "This is an example generated with a cloned voice.",
    "response_format": "mp3",
    "speed": 1
  }'

To adjust volume, speed, pitch, or emotion for the cloned voice, pass MiniMax-native options through metadata.voice_setting:

{
  "model": "minimax-speech-2.8-hd",
  "voice": "voice_abc123",
  "input": "This is a cloned voice example with emotion and volume settings.",
  "response_format": "mp3",
  "metadata": {
    "voice_setting": {
      "voice_id": "voice_abc123",
      "speed": 1,
      "vol": 1,
      "pitch": 0,
      "emotion": "happy"
    },
    "audio_setting": {
      "format": "mp3",
      "sample_rate": 32000,
      "bitrate": 128000,
      "channel": 1
    },
    "language_boost": "English"
  }
}

Keep voice and metadata.voice_setting.voice_id consistent. In /v1/audio/voice-cloning, model selects the cloning capability; minimax-voice-cloning is recommended. If you pass text to generate preview audio, set preview_model, such as speech-2.8-hd or speech-2.8-turbo.

Native Options Example

To set sample rate, volume, emotion, language boost, pronunciation dictionaries, or other MiniMax-native options, send them in metadata:

{
  "model": "minimax-speech-2.8-hd",
  "voice": "male-qn-qingse",
  "input": "Welcome to Cubicspaces MiniMax speech synthesis.",
  "response_format": "mp3",
  "metadata": {
    "audio_setting": {
      "format": "mp3",
      "sample_rate": 32000,
      "bitrate": 128000,
      "channel": 1
    },
    "voice_setting": {
      "voice_id": "male-qn-qingse",
      "speed": 1,
      "vol": 1,
      "pitch": 0,
      "emotion": "happy"
    },
    "language_boost": "English",
    "subtitle_enable": false
  }
}

Common native fields:

Field	Description
audio_setting.format	Audio format, such as mp3, pcm, flac, or wav
audio_setting.sample_rate	Sample rate, such as 8000, 16000, 22050, 24000, 32000, or 44100
audio_setting.bitrate	MP3 bitrate, such as 32000, 64000, 128000, or 256000
voice_setting.voice_id	Voice ID. For mixed voices, use it with timbre_weights
voice_setting.speed	Speech speed, commonly 0.5 to 2
voice_setting.vol	Volume, commonly (0, 10]
voice_setting.pitch	Pitch, commonly -12 to 12
voice_setting.emotion	Emotion, such as happy, sad, angry, fearful, disgusted, surprised, calm, fluent, or whisper
voice_modify	Voice pitch, timbre, intensity, and sound effect controls
pronunciation_dict	Pronunciation dictionary
subtitle_enable	Whether to request subtitle output. Actual behavior depends on model support

Notes

• This page covers synchronous speech generation. The response body is the generated audio content.
• metadata overrides extension fields with the same name. Do not change model inside metadata, to keep request behavior predictable.