Zero-shot voice cloning
Core model parameters ~0.5B
Model files 1.85 GB
Clone a voice from a 5 s recording, no training; it can speak other languages, dialects and emotions
- 5 s of audio
- Cross-lingual
- Dialects & emotion
Recorded on the board
Screen recording of the board web demo, recorded on Qualcomm QCS8550.
Highlights
Zero-shot voice cloning
A new voice from a 3-5 s recording and its transcript, made on the board, no training.
9 languages, cross-lingual
Chinese, English, Japanese, Korean, German, French, Spanish, Italian and Russian, in any voice.
Emotion and dialects
Happy, sad, angry, faster or slower, Sichuanese or Cantonese, set by an instruction.
Vocal tags
Laughter, breath and emphasis tags in the text for a more natural delivery.
Streaming, any length
Plays while generating; long text is split into segments and joined into one stream.
Fully on-device
Cloning runs on the CPU, synthesis on the NPU: no cloud, no GPU.
Pipeline
From input to output: where every step runs.
- 01Browser
Reference
A 3-5 s recording and its transcript, or a built-in voice
- 02CPU
Cloning
Speech tokenizer and speaker model make a voice pack
- 03NPU
Speech tokens
Text to speech tokens in the voice, following the instruction
- 04NPU
Waveform
Flow matching and a vocoder turn the tokens into a waveform
- 05Browser
Plays as generated
Each step of audio plays as it arrives, with a live waveform
Supported hardware
- Core model parameters
- ~0.5B
- Model files
- 1.85 GB
- License
- Pandora License