phoonnx-rapido

Per-architecture mono-repo of Rápido voices for phoonnx. One model per subfolder, each with a clean phoonnx config.json. Part of the phoonnx community mirror.

Rápido is a 21M-parameter flow-matching transformer over the VoxCPM2 AudioVAE decoder. It reads graphemes and writes 48 kHz audio. Each subfolder holds four ONNX graphs and config.json. model.onnx is the flow model. text_encoder.onnx, aligner.onnx and decoder.onnx are auxiliary graphs: the phoonnx voice index gives their URLs, and phoonnx downloads them with the voice.

subfolder	voice id	input	speakers	weights	license	source
edwixx__rapido-tts	edwixx/pt-BR-rapido	graphemes pt-BR	42	fp32	apache-2.0	edwixx/rapido-tts
edwixx__rapido-tts-int8	edwixx/pt-BR-rapido-int8	graphemes pt-BR	42	flow model dynamic int8, other graphs fp32	apache-2.0	edwixx/rapido-tts

Evaluation

Six pt-BR sentences, speaker 0, onnxruntime 1.31.1 on a laptop CPU with 4 threads. Word error rate from nemo-parakeet-tdt-0.6b-v3 (onnx-asr, int8).

weights	size	WER	real-time factor
fp32	251 MB	1/67	0.44
flow model int8	208 MB	1/67	0.40
all four graphs int8 (not published)	183 MB	2/67	0.55

The one error in the published voices is "quinze" transcribed as "15". With all four graphs in int8, the recogniser also drops "a" in "a cada", and the run is slower, because the decoder is mostly convolutions that dynamic quantization leaves in fp32.