MiniMax ASR transcription and Speech HD and Turbo synthesis.
16 integrated models · Official documentation ↗
MiniMax's current flagship TTS model (January 2026): 40 languages, seven emotions, voice cloning and inline sound tags such as (laughs), (sighs) and (breath) that add natural fillers and breathing.
The speed-focused tier of MiniMax's current Speech 2.8 generation (January 2026), with the same 40 languages and interjection sound tags as 2.8 HD at lower latency and price.
Speech 2.8 HD, MiniMax's newest voice model, used for asynchronous long-text jobs; it keeps the interjection tags that only the 2.8 models support and returns audio with sentence-level subtitles once the task finishes.
Asynchronous long-text synthesis on Speech 2.8 Turbo, pairing the newest generation's sound tags with the batch job API: up to 1M characters per task via file upload, with download links valid for nine hours.
MiniMax Speech 2.6 HD (October 2025), built for voice agents with a claimed sub-250 ms end-to-end latency, 40 languages and extra 'fluent' and 'whisper' emotions; MiniMax now lists it as legacy behind Speech 2.8 (January 2026).
Speech 2.6 Turbo (October 2025), the value tier of MiniMax's agent-oriented 2.6 generation, covering 40 languages at lower cost and latency than 2.6 HD; Speech 2.8 Turbo has replaced it as the current turbo model.
Long-form asynchronous rendering with Speech 2.6 HD, the October 2025 model that added 'fluent' and 'whisper' emotions; text goes in as a task (up to 50,000 characters, or 1M from a file) and audio is fetched later by file ID.
Speech 2.6 Turbo submitted through MiniMax's async T2A endpoint, a cheaper route for long narration where latency doesn't matter; output arrives as downloadable audio plus subtitle and metadata files.
MiniMax's second-generation HD TTS model (April 2025), focused on voice-clone similarity and audio quality across 24 languages with seven emotions; MiniMax now lists it as legacy behind Speech-2.6 and the current Speech-2.8.
The faster, cheaper sibling of Speech-02 HD from April 2025, tuned for rhythm, stability and low latency across 24 languages; superseded by Speech-2.6 Turbo and the current Speech-2.8 Turbo.
Speech-02 HD run through MiniMax's asynchronous long-text API: submit up to 50,000 characters inline or 1M via an uploaded file, then download the audio with sentence-level subtitles (links expire after nine hours).
Asynchronous batch interface for Speech-02 Turbo, aimed at audiobook-length text: tasks accept up to 1M characters from a txt or zip file and return audio, subtitle and metadata files rather than a stream.
MiniMax's text-to-image model (February 2025), with eight preset aspect ratios from 21:9 to 9:16, custom sizes up to 2048 px, up to nine images per request and prompts of up to 1,500 characters.
MiniMax's first-generation HD text-to-speech model (January 2025), launched with 17 languages, 300+ preset voices and voice cloning from about 10 seconds of audio; Speech-02, 2.6 and the current 2.8 have since superseded it.
MiniMax's speech-to-text model, covering 20 languages with automatic detection and Chinese-English code-switching; the API takes up to 500 seconds or 50 MB per request and offers SSE streaming, segment timestamps, diarization and SRT/VTT output.
Lower-latency, lower-cost turbo tier of MiniMax's first-generation Speech-01 TTS; it lacks the 2.6 and 2.8 models' wider language list and sound tags and is superseded by Speech-2.8 Turbo.
