Tasks and inputs
Qwen-Audio-3.1-TTS Flash is listed for Audio, Text-to-speech. It accepts text and produces audio. Use the publisher’s input format and preprocessing for your chosen deployment.
SyncDev model directory · Profile updated Nov 17, 2026
Qwen-Audio-3.1-TTS Flash is Alibaba's streaming text-to-speech model for voice assistants, live chat and customer service, with system and cloned voices.
The specification score summarizes recorded features and limits. It is not a benchmark result.
Pricing conditions: No published price found on Model Studio pages checked; confirm in the Model Studio console for your region.
Recorded modality support shows what this model can accept and produce. An undocumented modality is shown as unknown rather than assumed to be unsupported.
Accepts
Produces
A voice assistant lives or dies on the gap between the user finishing a sentence and the first syllable of the reply. qwen-audio-3.1-tts-flash is the member of Alibaba's 3.1 speech line built for that gap. It went live on Model Studio's international service on 12 October 2026, and its brief is narrow on purpose: real-time interaction, voice assistants, live conversation, customer service.
Requests go over plain HTTP POST to the SpeechSynthesizer path, authenticated with a DashScope API key. Two endpoints exist:
{WorkspaceId}.ap-southeast-1.maas.aliyuncs.com{WorkspaceId}.cn-beijing.maas.aliyuncs.comWithout streaming, the response returns a URL to the finished file, which you then download. That extra hop is fine for notifications and wasteful for conversation.
With the X-DashScope-SSE: enable header, audio arrives as server-sent events, each carrying a Base64 chunk. Play the first chunk as soon as it lands. Alibaba's own example asks for WAV at 24 kHz; other formats and rates are not listed on the guide we read, so test before you assume Opus or 16 kHz.
One limit of this interface: it expects the complete text in a single request. If your language model streams tokens and you want speech to start mid-sentence, Alibaba points you to its separate real-time synthesis interface instead.
Flash ships with a range of system voices; the documentation's example uses longanhuan_v3.1. You can also bring your own:
Language is not a separate parameter. You state it in the instruction field, which must match the language of the text. The same field takes natural-language direction on rate, emotion or style. Chinese dialects are available on voices that support them, either directly or through an instruction such as a request for Shanghainese.
Twenty-three control tags (moods like excited or serious, whispering, very slow or very fast speech) and seven nonverbal tags (a gasp, a sigh, a laugh, clearing the throat and others) can be written straight into the text. Control tags hold until the next one. For pronunciation of numbers and abbreviations, language_hints and hot_fix help where supported, and SSML handles pauses.
Pricing for this model was not on the Model Studio pages we read. Neither were maximum input length, a full format list or rate-limit figures. Check the console and the voice list in your region before committing.
Measure what your callers feel. Send fifty representative replies, streaming on, from the region you will deploy in. Record time to first audio chunk at p50 and p95, and listen for clipped starts. Then run the same script through Qwen-Audio-3.1-Realtime-Plus if you are weighing a full speech-to-speech design instead of a text-plus-TTS pipeline.
Other speech engines are collected under text-to-speech models.
Compare the model on the work you need to complete.
Qwen-Audio-3.1-TTS Flash is listed for Audio, Text-to-speech. It accepts text and produces audio. Use the publisher’s input format and preprocessing for your chosen deployment.
Compare pronunciation, voice consistency and time to the first playable audio. Check the service’s billing unit for the actual script.
Use the model identifier and account requirements for the selected API route. Confirm its request limits and output format with one complete task before connecting a larger workflow.
These records share a directory category. Check the source card, output type and deployment requirements before treating a model as a replacement.
zyphra
ZONOS3-Mini is Zyphra's open multilingual text-to-speech model: 800M MoE with ~100M active, zero-shot voice cloning in 43 languages and 80 ms first audio.
Alibaba
Qwen-Audio-3.1-TTS Plus is Alibaba's premium text-to-speech model for brand voices and dubbing, with fine-grained emotion tags and noise-robust voice cloning.
Mistral AI
Voxtral TTS 3 is Mistral's speech synthesis model for batch and streaming audio, with voice cloning, emotion tags and 35+ languages at $13 per 1M chars.
fish-audio
Fish Audio Drama 3 is an expressive 35B MoE text-to-speech model steered by bracketed directions, covering 83 languages, with downloadable weights.
heygen
HeyGen Voice is HeyGen's default text-to-speech model: about 39 languages, public and cloned voices, streaming with character timestamps.
heygen
HeyGen Voice Turbo is the low-latency version of HeyGen's TTS model for public voices and instant clones, at $4 per 1M characters.