Qwen's Voice Split: What You Can Rent (Audio-3.1) vs Run Locally (Qwen3-TTS/ASR)

lmoroney · x · 2026-10-10

Laurence Moroney untangles what people mean by "Qwen's new voice model" and lays out how to build a fully local voice loop.

The rented track: Qwen-Audio-3.1

Five hosted models — ASR, ASR-Next, TTS, TTS-Next and Realtime — all API-only, no open weights, no disclosed parameter counts.
Realtime is the headline: full-duplex streaming in and out, listens while it speaks and handles interruptions; supports function calling, web search, cloned voices and eight new system voices over WebSocket/WebRTC/AOQ; switches languages mid-call and reads emotion and intent; 262,144-token context, with sessions capped at 50 audio turns or 300 seconds of cumulative audio.
TTS-Next is billed as an "AudioGen" model producing speech, sound effects and ambience in one pass: up to 3,000 input characters, up to 240 seconds of podcast output, up to three 30-second reference clips, Chinese and English, non-streaming.
ASR covers 30 languages and 16 Chinese dialects.

The runnable track: Qwen3-TTS and Qwen3-ASR

Open-weight voice models released January 2026 under Apache-2.0, runnable on your own machine.
The post compares what each track does, what has been verified so far, and how to wire local TTS and ASR into a complete voice loop.

More from Infra