Qwen-Audio-3.1-TTS

// Overview

Qwen-Audio-3.1-TTS is the text-to-speech model in Qwen-Audio-3.1, the five-model speech release Alibaba's Qwen team announced on 23 September 2026 alongside Qwen-Audio-3.1-ASR, ASR-Next, TTS-Next and Realtime. In Alibaba Cloud Model Studio it is served as `qwen-audio-3.1-tts-flash`, a streaming model that Alibaba documents for real-time use such as voice assistants and customer service.

Alibaba's research page describes it as a production-oriented system built on a 12.5 Hz low-frame-rate speech tokenizer, which cuts autoregressive decoding cost, and a five-stage progressive training pipeline that pairs a language model with a flow-matching (FM) decoder and ends with reinforcement learning on both. The page lists 16 supported languages, seven of them added in this version, plus 20 Chinese dialect regions and one-pass synthesis of up to 3 minutes.

Control comes in two forms. Free-style natural-language instructions set role, emotion, speaking style, rate, timbre and accent, and 86 fine-grained inline tags adjust delivery at phrase and word level, including non-verbal events such as laughter, breathing, coughing and sighing. For voice cloning, Alibaba says the model stays robust when the reference recording is noisy, reverberant or unclear, without a separate denoising mode.

Alibaba's research page reports state-of-the-art results on many dimensions of SEED-TTS-Eval, CV3-Eval, instruction-following, long-form and robustness tests, and publishes CV3-Eval radar charts against MiniMax-Speech-2.8-HD, ElevenLabs-v3, Dots.TTS-2B (SOAR), VoxCPM2 and Qwen3-TTS-12Hz-1.7B-Base. It also states the model ranks first on the Artificial Analysis Text-to-Speech Leaderboard.

The release came with a price cut of about 70% for TTS, according to the Qwen launch post. Alibaba Cloud's price notice lists qwen-audio-3.1-tts-flash in the Singapore region at $0.23 per 1M input tokens and $1.87 per 1M output tokens from 22 September 2026 (UTC+8), replacing per-character billing of $0.15 per 10,000 characters.

// Benchmarks

// Pricing

Singapore region rates for qwen-audio-3.1-tts-flash from 22 September 2026 (UTC+8). The China (Beijing) model page lists 1.5 CNY input and 12 CNY output per 1M tokens.

Pricing source ↗

// Strengths

• 16 languages (seven new in 3.1) and 20 Chinese dialect regions in one model
• Natural-language instructions for role, emotion, style, rate, timbre and accent, plus 86 inline tags for word-level control and non-verbal sounds
• Voice cloning that Alibaba says holds up with noisy, reverberant or unclear reference audio
• One-pass long-form synthesis of up to 3 minutes
• Streaming output for real-time assistants and customer-service voices
• $0.23 / $1.87 per 1M input / output tokens in Singapore after the September 2026 price cut

// Best for

• Reach for it for voice assistants and customer-service agents that need streaming speech in several languages
• Reach for it for Chinese-market products that need dialect voices, such as Cantonese, Sichuan or Shanghai
• Reach for it for narration and dubbing directed by instruction rather than by picking a fixed voice
• Reach for Qwen-Audio-3.1-TTS-Next instead for podcasts, games and ads that need sound effects and ambience mixed with speech

// How to access

// Qwen-Audio (speech) — every version

The full lineage of the Qwen-Audio (speech) line, newest first. Every version has its own page — click any to compare specs, benchmarks and pricing.