Minimax

Minimax provides streaming TTS over WebSocket with multi-language support including English, Chinese, Japanese, and Korean. Vapi connects to Minimax via the speech-02-hd and speech-02-turbo model families.

Basic configuration

Set the voice on your assistant:

1
{
2
  "voice": {
3
    "provider": "minimax",
4
    "model": "speech-02-hd",
5
    "voiceId": "Wise_Woman"
6
  }
7
}

Subtitle timing for live captions (subtitleType)

Minimax can return subtitle data alongside synthesized audio, which Vapi forwards through the assistant.speechStarted client/server message. This is intended for live caption UIs and karaoke-style word highlighting.

1
{
2
  "voice": {
3
    "provider": "minimax",
4
    "model": "speech-02-hd",
5
    "voiceId": "Wise_Woman",
6
    "subtitleType": "word"
7
  }
8
}

You also need to subscribe to the message itself by adding "assistant.speechStarted" to your assistant’s clientMessages and/or serverMessages arrays.

How the timing actually works (and what it can’t do)

This is the most important part to understand before building on top of it.

Minimax synthesizes audio incrementally, but it only attaches subtitle metadata to the final audio chunk of each synthesis segment. Vapi streams every audio chunk to the call as soon as it arrives, but the wordsSpoken cursor only advances when that final chunk is reached. In practice, this means:

• You will receive one assistant.speechStarted event per Minimax synthesis segment, not one per word.
• That event fires near the end of the segment’s audio playback, not at the start. The wordsSpoken value jumps forward in segment-sized increments rather than ticking word by word.
• The timing.words[] array in each event carries the per-word start/end timestamps for the segment that just finished. You can use it to animate that segment retroactively, or to extrapolate forward during the next segment — but you cannot use it to drive smooth real-time highlighting in the current segment.
• Per-word timestamps are relative to the segment’s start, not the start of the call.

If your use case requires word-by-word highlighting at audio playback cadence with no interpolation, use ElevenLabs — its word-alignment timing arrives every 50–200ms with real per-word timestamps from the provider. Minimax word-progress is best suited to:

• Caption blocks that update once per spoken sentence/clause.
• “How far through the response are we” progress indicators.
• Post-hoc transcript annotation with word-level timing.

Other behaviors to be aware of

• totalWords === 0 is a valid value on the first event of a turn, before Minimax has confirmed the word count. Guard against divide-by-zero when computing progress fractions.
• force-say events (your firstMessage, queued say actions) are emitted as text-only events — no timing object — even when subtitleType: "word" is configured. This is because Minimax does not return subtitle metadata for these utterances.
• On user barge-in, no further events fire for the interrupted turn. The most recent wordsSpoken tells you how much of text was actually spoken before the interruption.
• CJK languages (Chinese, Japanese, Korean) are word-counted per ideograph/kana/hangul. A 30-character Japanese sentence reports totalWords: 30.

For the full event schema and timing shapes across all voice providers, see Server events → Assistant Speech Started.