Qwen3-TTS speech synthesis

About

Qwen3-TTS reads your script aloud. It ships on Sogni as three checkpoints under one family: nine built-in studio voices, a voice cloned from a few seconds of a real recording, and a voice invented from a written description. All three take the same script, the same languages and the same price.

The single most important thing to know is what the prompt is. It is not a description of the audio you want — it is the audio. Whatever you type is spoken character for character, so “a calm woman reading the evening news” gets read out as those seven words. The description of the voice belongs in a separate field; the prompt is only ever the words to be said.

Studio Voices is the one to reach for when nobody in particular has to be imitated: narration, voiceover, a line of dialogue. Nine voices ship with it — Serena, Vivian, Uncle Fu, Ryan, Aiden, Ono Anna, Sohee, Eric and Dylan. The nationality attached to each one describes the accent it carries, not the language it can read. On top of the voice you can add a written direction — “whispering, close to the mic”, “like a sports commentator” — which restyles the delivery without changing who is speaking.

Voice Clone reproduces a specific person from three to thirty seconds of them talking. Supply the reference recording’s exact transcript alongside it whenever you know it: with the transcript the model conditions on the recording itself, and without it it has only a speaker fingerprint and its own prior to fall back on. That difference is not subtle. In our own testing, cloning an accented speaker without the transcript produced a fluent General American voice that had lost the accent entirely — recognisably not her.

Voice Design invents a speaker who does not exist. Here the description is required and is doing all the work: age, gender, accent, timbre, pace, mood, the room it was recorded in. Describe one coherent person — contradictory directions produce a voice that shifts mid-sentence.

Language is set to auto by default, which infers it from the script and is the only setting that reads a code-switched line correctly. Pin a language only when auto mis-reads a name or a loanword, and be careful about pinning it on a cloned voice: forcing English onto a non-native speaker is one of the ways to lose the accent you were cloning for.

Speech is priced per 1,000 characters of script — the unit every TTS provider bills in, because the script rather than a requested duration decides how long the model runs. That is $0.42 for roughly 74 seconds of speech, with every take billed at a floor of 170 characters so a one-line take still covers loading the model. Generation is slower than real time: expect around 1.6 seconds of GPU for each second of audio.

Speak a line in the Sogni app under Speech, or call any of the three model ids from the JavaScript or Python SDK. It runs on the Sogni Supernet, a decentralized network of creator GPUs, so nothing is installed and no card is rented.

Three modes, three checkpoints

These are three separate checkpoints rather than three presets of one, so each takes its own inputs. Everything else — script length, languages, creativity, takes, output format and price — is identical across all three.

Prompting tips

The prompt is the script, so writing a good one is writing for the ear rather than for the eye:

• Write the words, not the voice — Anything in the script is read out loud. “Warm female narrator, 30s” belongs in the style direction or the voice description — put it in the script and you will hear it spoken.
• Punctuation is prosody — Full stops, commas, question marks and ellipses are what control pauses and intonation. Keep them. Line breaks are not pauses and do nothing.
• Spell out anything ambiguous — Write numbers, dates, currency and abbreviations the way they should be said — “nineteen eighty-four”, “twelve dollars fifty”, “Doctor Chen” — rather than leaving the model to guess.
• Leave the language on auto — Auto infers the language from the script and is the only setting that handles a line that switches between two of them. Pin one only when auto mis-reads a specific name or loanword.
• Give the clone a transcript — Type out exactly what the reference recording says. It is the difference between a clone that keeps the speaker’s accent and one that reads your script in a generic voice.
• Split a long script — The limit is 4,096 characters and a longer script is refused rather than truncated. Break it at a natural pause and run the parts as separate takes.
• Re-roll rather than rewrite — Delivery varies between takes of the same script. Ask for several takes and pick, before you start editing words that were never the problem.

What each mode needs from you

The script, the language, the creativity and the number of takes are the same everywhere. What differs is how the voice gets decided.

A written direction restyles a studio voice without changing who it is; on Voice Design the same field defines the identity. Voice Clone takes no direction at all — the recording decides.

The nine studio voices

Every one of them speaks all ten supported languages. The nationality below is the accent the timbre carries, not a limit on what it can read.

The rest of the request

Cloning a voice

The transcript is optional and you should almost always send it. Without it the clone has a speaker fingerprint and its own prior; with it, it conditions on the recording.

const project = await client.projects.create({
  type: 'audio',
  modelId: 'qwen3_tts_1.7b_voice_clone_bf16',
  positivePrompt: 'The compute is borrowed. The voice is yours.',
  numberOfMedia: 1,
  referenceAudio: readFileSync('reference.wav'),
  // The exact words spoken in reference.wav. This is what keeps the accent.
  referenceText: 'okay say goodnight daddy i love you so i say goodnight to you',
  outputFormat: 'wav',
});

Pricing

Use pay-as-you-go Spark packs for each take (1 Spark = $0.005), or choose a flat-rate Sogni plan for credit-free fair-use generation in the app.

Priced per 1,000 characters of script, with a 170-character floor on every take. The script rather than a requested duration is what decides how long the model runs, so characters are the unit — the same one every other TTS provider bills in.

1 Spark = $0.005. Pay as you go with Spark packs, or speak under fair use on a flat monthly Sogni plan.

API

One Sogni API key reaches every model on the Supernet — call Qwen3-TTS with the exact model id for the variant you want.

import { SogniClient } from '@sogni-ai/sogni-client';

const client = await SogniClient.createInstance({
  appId: crypto.randomUUID(),
  apiKey: process.env.SOGNI_API_KEY,
  network: 'fast',
});

const project = await client.projects.create({
  type: 'audio',
  modelId: 'qwen3_tts_1.7b_custom_voice_bf16',
  // The prompt IS the script — every character of it is read aloud.
  positivePrompt: 'Every render on the Sogni Supernet runs on a graphics card somebody else owns. Including this sentence.',
  numberOfMedia: 1,
  speaker: 'serena',
  instruct: 'warm and unhurried, close to the mic',
  outputFormat: 'mp3',
});

const [url] = await project.waitForCompletion();
console.log(url); // audio link — download within 24h

curl https://api.sogni.ai/v1/creative-agent/workflows \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer $SOGNI_API_KEY" \
  -d '{
    "input": {
      "title": "First render",
      "steps": [{
        "id": "step1",
        "toolName": "generate_speech",
        "arguments": { "prompt": "Every render on the Sogni Supernet runs on a graphics card somebody else owns. Including this sentence.", "model": "voice", "voice": "serena" }
      }]
    },
    "confirm_cost": true
  }'

Example uses qwen3_tts_1.7b_custom_voice_bf16; use any exact model id from the table above. The same three ids work from the Python SDK, and the REST workflow API reaches them through generate_speech, where model selects the mode — voice, clone or design. Full reference at docs.sogni.ai.