Qwen3-TTS — Speech AI Model

Details

• CategorySpeech
• Year2026
• LicenseApache-2.0

Compliance & Provenance

• ProviderAlibaba (open weights) · Specialized
• EU AI Act RiskLimited Risk
• Art. 50 TransparencyRequired — AI-generated outputs are marked

Inputs & Outputs

TextInput · string

Text to speak (the narration script)

SoundOutput · sound

Synthesized voiceover audio (WAV)

Parameters

voicestring · default Default · options: Default, Sohee, Ryan, Aiden, Ono Anna, Vivian, Serena, +3 more

Narrator voice preset.

Qwen3-TTS CustomVoice speaker. 'Default' picks the native speaker for the spoken language (Korean or Hangul text -> Sohee, English -> Ryan). Native languages: Sohee Korean; Ryan, Aiden English; Ono Anna Japanese; Vivian, Serena, Uncle Fu Chinese; Dylan Beijing dialect; Eric Sichuan dialect. A speaker reading a non-native language keeps an accent.

languagestring · default Auto · options: Auto, Korean, English · advanced

Spoken language hint (Auto follows the text).

speedfloat · default 1 · range 0.5–2

Speaking speed. 1.0 = as generated; >1 faster, <1 slower. Pitch is kept.

Playback speed of the synthesized voice, 0.5-2.0 (default 1.0). The audio is re-timed after synthesis with its pitch kept, so 1.1 means exactly 10% faster. Values outside the range are clamped.

pronunciationsstring · default · advanced

Respell words the voice misreads: written=spoken, separated by ; or new lines, e.g. cnaps.ai=see-naps dot A I.

Pronunciation rules applied only to the text sent to the voice model: written=spoken entries separated by ';' or new lines, e.g. `cnaps.ai=see-naps dot A I; cnaps=see-naps`. Matching is case-insensitive and whole-word: an English letter or digit directly before or after a written form that starts or ends with one blocks the match (cnaps never fires inside cnapsule), and a written form starting with Hangul must not follow another letter (a rule for 냅스 does not touch 씨냅스). Korean particles may follow directly: cnaps.ai에서 and 씨냅스는 both match. Longer written forms win over shorter ones and replaced text is not re-scanned. Max 50 rules and 2000 characters; a rule without '=' or with an empty side fails the run. Captions should come from the original script (feed the same script to Whisper's Script port), not the respelled text.

loudnessstring · default off · options: off, -14 LUFS, -16 LUFS, -20 LUFS · advanced

Normalize the voice to a fixed loudness. off keeps the level as generated.

Loudness normalization after synthesis and the speed change (ffmpeg loudnorm, single pass, true peak -1.5 dBTP). 'off' keeps the level as generated; '-14 LUFS' is loud (social/streaming), '-16 LUFS' the usual voiceover level, '-20 LUFS' quieter, leaving headroom for music. Default off, so existing flows keep their level; turn it on when the level must match other clips.

Tags

• text-to-speech
• tts
• voiceover
• narration
• audio
• speech
• qwen3

Alternatives in Audio Models

• Music (Lyria)

  Google Lyria: text (and optional image) to music — full songs with vocals and lyrics, or 30-second clips. Needs API key.

• Text to Music (ACE-Step 1.5)

  ACE-Step 1.5: generate music from a style prompt, with optional lyrics, key, tempo and time signature. 48 kHz stereo, 5-600s.

• Text to Sound Effect (MOSS-SoundEffect v2)

  MOSS-SoundEffect v2.0: generate a sound effect from an English description. 48 kHz mono, 0.5-30s.

• Text to Speech (Gemini)

  Google Gemini TTS: turn text into a spoken voiceover (Korean/English). Needs API key.

• Text to Speech (Qwen3 Voice Design)

  Qwen3-TTS VoiceDesign: on-device text-to-speech with a voice you describe in words. No API key.

Resources

• huggingface