開放模型與邊緣 AI
Inflect v2 Updates Ultra-Compact Local TTS Toolkit, Letting Users Adapt Voices and Export a 4M-Parameter Model to ONNX
Inflect-Nano-v2 keeps the complete text-to-waveform model to 3.97M parameters, and a recent update adds a runnable single-speaker adaptation workflow. It is suited to offline and edge voice interfaces, but its training data, cross-lingual performance, and quality comparisons still lack independent verification.
Independent developer Owen Song has updated Inflect v2, packaging two fixed-voice English TTS models and adaptation tools into a ready-to-run local toolkit. Inflect-Nano-v2 has 3,966,721 deployable parameters, with FP32 weights totaling 15.97 MB; the quality-focused Micro-v2 has 9,356,513 parameters and weighs 37.53 MB. Both include a 24 kHz neural waveform decoder—rather than counting only the text encoder—and can run on CPU or CUDA through the same Python API.
The architecture is a parameter-efficient member of the VITS family, combining an English phoneme front end, monotonic alignment, stochastic latent-variable synthesis, residual coupling flows, and an aliasing-reduced waveform decoder. Nano uses a 128-dimensional latent representation, a three-layer, two-head text encoder, and four flow coupling blocks. The recently added adaptation toolkit can import single-speaker recordings with sentence-level transcripts, warm-start from a Nano or Micro checkpoint, resume interrupted training, retain validation audio, and ultimately export a PyTorch or ONNX model. This moves Inflect beyond fixed-voice weights and turns it into a foundational tool for creating private voices for embedded products.
For phones, Raspberry Pi devices, games, and interfaces that cannot upload audio to the cloud, a complete model in the tens-of-megabytes range can reduce memory use, download size, and cold-start costs. A fixed seed also helps with regression testing. However, phonemization still depends on eSpeak-ng, which is not included in the parameter count. The public models have only been validated with fixed English voices; adapting them to languages such as Chinese requires new phonemization rules and substantial paired data, so this should not be treated as few-shot voice cloning. The training-corpus construction process has also not been fully disclosed. Before adoption, teams should independently test long-form prosody, numbers and proper nouns, the CPU real-time factor, and distortion introduced by quantization or ONNX export.