Chatterbox Turbo — 350M Zero-shot TTS (Resemble AI)

TL;DR

Resemble AI’s Chatterbox Turbo is a 350M-parameter open-source zero-shot TTS model paired with a distilled audio diffusion decoder that drops sampling from 10 steps to 1. The demo page positions it as a lightweight voice-agent model (low VRAM, low latency) that still supports inline paralinguistic tags — [laugh], [cough], [chuckle], [whisper], [shush] — and creative narration. Samples on the page cover IVR menus, customer support, character voices (C-3PO, Dwight Schrute, Frasier, Snoop), and emotive monologue. Resemble pitches it as the open counterpart to their managed sub-200ms-latency hosted TTS service.

Key claims

350M-parameter architecture, explicitly framed as smaller-VRAM / lower-compute than larger open TTS systems [§350M Zero-shot TTS].
A new distilled audio diffusion decoder reduces generation from 10 steps to 1 while retaining high fidelity [§350M Zero-shot TTS].
Native paralinguistic tag support: [laugh], [cough], [chuckle], plus [whisper] and [shush] observed in the samples [§350M Zero-shot TTS, sample transcripts].
Targeted at low-latency voice agents but demonstrated on creative narration / character voices [§350M Zero-shot TTS, §Voice Agents samples].
For sub-200ms production latency the page directs to Resemble’s managed service rather than the open model [§350M Zero-shot TTS footer].

Method

The demo page is a samples-only release: no architecture diagram, no benchmark numbers, no training-data description. The two mechanical claims are (a) ~350M parameters total and (b) a distilled audio diffusion decoder running at a single sampling step (down from a 10-step teacher). The voice-agent samples include inline bracket tags for non-lexical events (laughs, chuckles, shushes) that the model renders as the appropriate paralinguistic sound at the tagged position. Zero-shot voice cloning is implied by the “Zero-shot TTS” framing but no reference-audio protocol is described on the page.

Results

No quantitative benchmarks on the demo page — only audio samples. Sample categories shown: IVR menus (CareLine Medical, Support Services), customer-support agents, satirical sales pitches (Jerry Seinfeld, Dwight Schrute), pop-culture impressions (C-3PO, Frasier Crane, Snoop Dogg), and emotive monologue (excited “besties” voice, menacing villain). Quality has to be judged by listening; the Luma share note is “samples seem pretty good”.

Why it’s interesting

Sits in the same band as Fish Audio S2 Pro — SOTA Open-Source TTS (Dual-AR + GRPO) — both are 2026 open zero-shot TTS releases targeting voice-agent latency — but Chatterbox Turbo is an order of magnitude smaller (350M vs Fish S2 Pro’s 4B + 400M Dual-AR) and bets on distilling the audio decoder down to 1 sampling step rather than on scaling an autoregressive codec model. That puts the diffusion-distillation lever from Diffusion Distillation (so far applied mostly to image/video teachers like Wan2.1, SD3.5-M, FLUX) onto a TTS decoder, where the 10→1 step claim is the load-bearing efficiency move. The inline paralinguistic-tag UI ([laugh], [chuckle], [shush]) is the same convergence direction as Gemini 3.1 Flash TTS: Google text-to-speech with scene direction and audio tags (announcement) (Gemini 3.1 Flash TTS scene tags), Qwen3-TTS (version 2025-11-27) announcement — 49 voices, 10 languages, 8 Chinese dialects, ~97ms first-packet latency (Qwen3-TTS), and Fish Audio S2 Pro — SOTA Open-Source TTS (Dual-AR + GRPO)‘s 15,000+ tag interface — natural-language tags as the control surface for TTS. Contrast with CosyVoice 2: Scalable Streaming Speech Synthesis with Large Language Models which uses a flow-matching mel decoder and a multi-step diffusion sampler; Chatterbox Turbo collapses that to a single step.