Minimax speech 2.8 brings human warmth to AI voices with native filler words and high-fidelity cloning

MiniMax dropped Speech 2.8 today, the latest version of its speech synthesis engine. The big goal here: kill the flat, robot-like sound that has haunted AI voices for years. This update goes after four weak spots: modeling filler words naturally, cloning voice timbres with high fidelity, producing noise-free audio, and handling cross-language synthesis more smoothly.

The 'too perfect' problem

The company fingered the core issue: AI voices sound “too perfect.” Real human speech is messy. It's filled with breaths, pauses, and hesitations, little signals that carry emotion and emphasis. Speech 2.8 now models filler words like “um,” “uh,” and “ah” natively, keeping their natural timing and rhythm intact.

In one demo audio MiniMax shared, the model produces a conversational monologue. You hear an intake of breath, natural pauses, and spontaneous-sounding fillers. The script ends with a twist: the speaker reveals they are not human but the Speech 2.8 model itself, urging listeners to “pay close attention to the breathing, the tone, and even the casual rambling.”

10-second voice cloning

Timbre cloning got a big upgrade, too. Speech 2.8 needs only 10 seconds of original audio to lock onto what MiniMax calls a person's “voice fingerprint”, not just tone, but also breath patterns, quality, and idiosyncratic speaking rhythm.

MiniMax offered a granular breakdown of one cloning example. It highlighted three features: a smooth “mature texture” with low-frequency chest resonance and nasal characteristics; a “lazy yet natural breathing style” with deep breaths and casual exhales that make short phrases feel like real interaction; and a micro-sarcastic speech pattern, including a slow-start-fast-end rhythm and a trailing intonation at sentence ends that conveys playful skepticism.

Studio-grade audio clarity

The audio processing pipeline got a full overhaul to scrub out background noise and digital artifacts. Output is now described as “clean and transparent,” comparable to a recording from a professional studio. A demo clip drives the point home with a poetic narration about a forest morning, complete with subtle wind and leaf sounds that blend seamlessly with the synthetic voice.

Cross-language improvements

Speech 2.8 takes on a persistent headache in multilingual speech synthesis: accent leakage. The first targeted fix improves Japanese synthesis from Chinese timbres, eliminating previous phoneme misalignment and unnatural intonation. MiniMax says the model now supports 40+ languages from a single voice profile, with more language pairs slated for optimization in future updates.

A Chinese-Japanese bilingual demo shows the model switching between the two languages mid-sentence, maintaining consistent voice quality and natural prosody in both.

Availability

Speech 2.8 is live now through the MiniMax open platform and the MiniMax Audio product. The company slots the model into its broader mission: “Intelligence with Everyone.”