Qwen3-TTS
A text-to-speech model based on the Qwen3 large language model. It is designed for high accuracy, low latency, and low cost, with applicable use cases including voice assistants, audiobooks, and accessibility tools.
Timeline (last 90 days)
The reference audio should be clean single-speaker speech with no background music and no reverb, otherwise zero-shot voice cloning will introduce noise.
The Qwen3-TTS Base model used in the article does not support instruction control (such as specifying emotion, tone, or speaking rate); this capability requires the Qwen 1.7B CustomVoice or VoiceDesign model, native runtime support, and the Dart API to all support it simultaneously.
On M4 Max, Omni 3B has a median dialogue generation time of 171 ms and a peak process RSS of 2.57 GiB; Qwen3.5 0.8B has a median dialogue generation time of 96 ms and a peak process RSS of 1.08 GiB; Qwen3-TTS 1.7B has a synthesis time of 1.78 seconds and a peak process RSS of 3.85 GiB.
On M4 Max, the text route using Omni 3B + TTS 1.7B (4.61 GB) has a median full-turn latency of 4.27 seconds, with a peak process RSS of 3.86 GiB.
The small model pair halves the download size and reduces peak memory by about 35% compared with the original model pair, but in this trial the full turn was actually slower.
TTS zero-shot voice cloning does not train a new model; instead, it encodes the reference voice into a speaker embedding, which is injected as a condition during synthesis, so the same set of weights can provide different timbres for different characters.
Nari has launched two speech models based on Qwen3: a text-to-speech (TTS) model and an automatic speech recognition (ASR) model
Qwen-Audio-3.0-TTS支持在文本中嵌入内联标签控制语音，示例包括[whisper]、[angry]、[breaths]、[laughs]
All Facts (8)
The reference audio should be clean single-speaker speech with no background music and no reverb, otherwise zero-shot voice cloning will introduce noise.
50%UnverifiedThe Qwen3-TTS Base model used in the article does not support instruction control (such as specifying emotion, tone, or speaking rate); this capability requires the Qwen 1.7B CustomVoice or VoiceDesign model, native runtime support, and the Dart API to all support it simultaneously.
50%UnverifiedOn M4 Max, Omni 3B has a median dialogue generation time of 171 ms and a peak process RSS of 2.57 GiB; Qwen3.5 0.8B has a median dialogue generation time of 96 ms and a peak process RSS of 1.08 GiB; Qwen3-TTS 1.7B has a synthesis time of 1.78 seconds and a peak process RSS of 3.85 GiB.
50%UnverifiedOn M4 Max, the text route using Omni 3B + TTS 1.7B (4.61 GB) has a median full-turn latency of 4.27 seconds, with a peak process RSS of 3.86 GiB.
50%UnverifiedThe small model pair halves the download size and reduces peak memory by about 35% compared with the original model pair, but in this trial the full turn was actually slower.
50%UnverifiedTTS zero-shot voice cloning does not train a new model; instead, it encodes the reference voice into a speaker embedding, which is injected as a condition during synthesis, so the same set of weights can provide different timbres for different characters.
50%UnverifiedQwen-Audio-3.0-TTS支持在文本中嵌入内联标签控制语音，示例包括[whisper]、[angry]、[breaths]、[laughs]
50%UnverifiedNari has launched two speech models based on Qwen3: a text-to-speech (TTS) model and an automatic speech recognition (ASR) model
60%