MiniMax Speech 2.8: Text-to-Speech with Native Sound Tags, 10-Second Voice Cloning, API Options and Pricing

What MiniMax Speech 2.8 is

MiniMax Speech 2.8 is a text-to-speech model series from MiniMax, announced on 23 January 2026 and listed in MiniMax's release notes with the summary natural sound tags, lifelike voice and pristine audio quality. It comes in two versions on the MiniMax API: speech-2.8-hd and speech-2.8-turbo, which MiniMax describes as having the same capabilities as hd, optimised for speed and cost. MiniMax says its aim is synthetic speech that feels truly human. It is a proprietary model with no published weights, offered through the MiniMax Open Platform and the MiniMax Audio app. On LMArena it has no LMArena rating.

Native sound tags

MiniMax's central change in Speech 2.8 is native sound tags. MiniMax says earlier AI voices felt cold because they were too perfect, while real speech carries breaths, pauses and hesitations. Speech 2.8 models colloquial fillers such as um, uh and ah to keep the rhythm, pitch and pauses of human dialogue. In the API, interjection tags placed in the text work only with speech-2.8-hd and speech-2.8-turbo. The supported tags include (laughs), (chuckle), (coughs), (clear-throat), (groans), (breath), (pant), (inhale), (exhale), (gasps), (sniffs), (sighs) and (snorts).

Cloning, clarity and languages

MiniMax says it optimised feature extraction for voice cloning, so that a 10-second sample is enough for Speech 2.8 to capture a speaker's texture, breathiness and speaking pace. It also re-engineered its processing to remove background noise and synthetic distortion, aiming for studio-like clarity. For cross-lingual speech, MiniMax says it reduced accent bleed, starting with the Mandarin-Japanese pair, where it fixed unnatural tones and pronunciation shifts, and said more language pairs would follow. Its English demo shows a narrator voice with natural fillers and rhythmic pauses.

Using the API

MiniMax offers Speech 2.8 through a synchronous T2A endpoint for real-time synthesis, with volume, pitch, speed and mixing controls and bitrate and sample-rate options, which MiniMax suggests for short-form conversational audio. A synchronous request must be under 10,000 characters. An asynchronous T2A endpoint handles long-form synthesis of up to 1 million characters per request, retrieved later. Non-streaming output can be MP3, WAV or FLAC and streaming output is MP3. The API also supports pronunciation dictionaries, a language boost setting for minority languages and dialects, voice effects, subtitles and a quality text-normalisation mode for non-streaming requests. MiniMax's sample request uses the voice English_expressive_narrator with speed, volume and pitch settings, a 32 kHz sample rate, 128 kbps MP3 output and a spacious echo sound effect, and the response reports the audio length, size and the number of characters billed.

Pricing

On the MiniMax API, speech-2.8-hd costs $100 per million characters and speech-2.8-turbo costs $60 per million characters, for both synchronous and asynchronous synthesis. MiniMax also sells Speech 2.8 access through MiniMax Audio packages. The same API still lists the earlier speech-2.6, speech-02 and speech-01 models in hd and turbo versions, though the interjection tags are limited to the 2.8 models.

Official Resources