How neuphonic/neutts works
NeuTTS is an on-device text-to-speech library built around language-model backbones and neural audio codecs. The main workflow accepts text and a speaker reference, prepares model input, generates speech tokens, decodes them to audio, and returns a waveform; GGUF models also support streaming. The graph preserves the Air and emotional 2E variants, multilingual phonemization, cloning, and watermarking. Some internal inference details and codec wiring are only partially excerpted, so edges are limited to documented or visible relationships.
Main entry points: neutts/neutts.py NeuTTS API · neutts_2e.py NeuTTS 2E · neuttsair/neutts.py NeuTTS Air · phonemizers.py Phonemization
Components 12
TTS interfaces
- NeuTTS API
neutts/neutts.py - NeuTTS Air
neuttsair/neutts.py - NeuTTS 2E
neutts/neutts_2e.py
Speech generation
- Speech backbone
neutts/neutts.py - Audio codec
neutts/neutts.py - Audio watermark
neutts/neutts.py - Streaming output
neutts/neutts.py
Language and voices
- Speaker references
neutts/neutts_2e.py - Phonemization
neutts/phonemizers.py
Other components
- Application user
- Text and reference
- Generated waveform
Connections 14
- Application user → NeuTTS API: uses
README.md - Text and reference → NeuTTS API: provides input
README.md - Application user → NeuTTS Air: uses
README.md - Application user → NeuTTS 2E: uses
README.md - NeuTTS Air → NeuTTS API: inherits
neuttsair/neutts.py - NeuTTS 2E → NeuTTS API: delegates inference
neutts/neutts_2e.py - NeuTTS 2E → Speaker references: loads references
neutts/neutts_2e.py - NeuTTS API → Phonemization: uses
neutts/neutts.py - NeuTTS API → Speech backbone: generates tokens
neutts/neutts.py - NeuTTS API → Audio codec: decodes tokens
neutts/neutts.py - NeuTTS API → Audio watermark: watermarks audio
neutts/neutts.py - NeuTTS API → Streaming output: yields chunks
neutts/neutts.py - NeuTTS API → Generated waveform: returns waveform
neutts/neutts.py - Streaming output → Generated waveform: yields audio
neutts/neutts.py