Qwen News 07/20/2026 AI Rating: Medium

Alibaba's Tongyi Lab Unveils "Qwen-Audio-3.0-TTS" — Claims Leaderboard Top Spot at a Third of ElevenLabs' Price

#Qwen#Alibaba#TTS#Text-to-Speech

On July 20, 2026, Alibaba’s Tongyi Lab unveiled a new text-to-speech (TTS) model called “Qwen-Audio-3.0-TTS.” Rather than shipping as open weights, it’s offered as a hosted API through Alibaba Cloud Model Studio in two tiers — “Flash” and “Plus” — covering 16 languages. The Plus tier reportedly took the top spot on Artificial Analysis’s text-to-speech leaderboard, and pricing is set at roughly a third of what competitors like ElevenLabs charge.

Details

  • Two tiers: “Flash” (model ID qwen-audio-3.0-tts-flash) is tuned for real-time conversation with first-packet latency in the 300ms range, while “Plus” (qwen-audio-3.0-tts-plus) prioritizes naturalness and timbre fidelity over raw speed
  • Pricing: $27.59 per 1 million characters — roughly a third of what rivals such as ElevenLabs and MiniMax charge
  • Language support: 16 languages including Arabic, Chinese, English, French, German, Indonesian, Italian, Japanese, Korean, Malay, Portuguese, Russian, Spanish, Tagalog, Thai, and Vietnamese, plus 20 Chinese dialect regions
  • Benchmark results: Plus ranks #1 on the Artificial Analysis Text-to-Speech leaderboard with an Elo rating of roughly 1,236, and posts the best word/character error rate (WER/CER) in 10 of the 16 languages. Average WER/CER is 3.87 for Flash and 3.96 for Plus
  • Speaker similarity: Plus averages 82.75 and Flash 80.44 across all languages, a claim of strong voice fidelity
  • Throughput trade-off: Plus generates only around 16 characters per second, notably slower than competitors like Simba 3.2 (30.2 chars/sec), Gemini 3.1 Flash TTS (27 chars/sec), and Sonic 3.5 (120 chars/sec)
  • Delivery model: API-only, not open-weight. Available via bidirectional WebSocket streaming from Singapore and Beijing regions

How to try it

  • Accessible via the DashScope SDK through Alibaba Cloud Model Studio, with sample code provided in Python, Java, Go, C#, PHP, and Node.js
  • No open-weight release has been announced; usage currently requires going through the hosted API
Source https://www.marktechpost.com/2026/07/20/alibabas-tongyi-lab-releases-qwen-audio-3-0-tts-a-hosted-text-to-speech-model-in-flash-and-plus-tiers-across-16-languages/

Related Articles

More Qwen articles →
Qwen News 08/13/2026 Read AI Rating: High

Qwen3.8-Max Open Weights Ship — Under a New Revenue-Sharing License, Not Apache 2.0

Alibaba released the open weights of Qwen3.8-Max, its first Max-class model to be open-sourced, on Hugging Face and ModelScope — but under a bespoke license requiring large commercial users to obtain a separate agreement, breaking from Qwen's usual Apache 2.0 terms.
Qwen News 08/03/2026 Read AI Rating: High

Alibaba Drops the "Preview" Label From Qwen3.8-Max, Open Weights Due Next Week

Alibaba has made Qwen3.8-Max widely accessible via Model Studio APIs, dropping its preview status ahead of an open-weight release planned for the following week, alongside a public beta of the QwenWork agent platform.
Qwen News 07/30/2026 Read AI Rating: Medium

Alibaba's Tongyi Lab Launches "Qwen-Audio-3.0-ASR" Speech Recognition Models — Covers 7 Chinese Dialect Groups and 30 Languages

Alibaba's Tongyi Lab has released a new automatic speech recognition (ASR) lineup, "Qwen-Audio-3.0-ASR," as a hosted API. It ships streaming and file-transcription variants covering China's seven major dialect groups, 20+ regional accents, and 30 languages overall.
Qwen News 07/25/2026 Read AI Rating: Medium

Alibaba Launches Qwen3.7-Flash, a Lightweight, Low-Cost Vision-Language Model with Stronger Agent Skills

Alibaba has released Qwen3.7-Flash, the lightweight tier of its Qwen3.7 series, delivering a comprehensive upgrade in multimodal understanding and agent execution over the prior 3.6-Flash, with a 1M-token context window and low pricing.