metadata
block_size: 2048
sample_rate: 44100
latent_size: 12
vocoder: 042-jvs-100m-xfermulti_0abe2b072b_streaming_norm.ts
dataset: John Van Stan (LibriTTS)
vocoder_type: RAVE
alignment_type: DCA
likelihood_type: NSF
text_encoder_type: CANINE
tungnaa_116_jvs
dimensions
block size: 2048
sample rate: 44100
latent size: 12
dataset
JVS (Hi-Fi TTS speaker 9017)
vocoder
models/vocoder/042-jvs-100m-xfermulti_0abe2b072b_streaming_norm.ts
training
tungnaa commit 09ecdcd532eac3d454a8b4e28e896bca5bccbf9f
tungnaa trainer --experiment 117-jvs-e2emulti-mask-ends --model-dir /data/users/victor/ivoice-models --log-dir /data/users/victor/ivoice-logs --manifest /data/users/victor/tmp/ivoice_prep_100m_0abe_multi/9017_manifest_clean_train.json --rave-model /data/users/victor/rave-v2/runs/042-jvs-100m-xfermulti_0abe2b072b/version_0/checkpoints/042-jvs-100m-xfermulti_0abe2b072b_streaming_norm.ts --lr 3e-4 --lr-text 3e-5 --epoch-size 200 --save-epochs 20 --device cuda:0 train
notes
trained with full JVS dataset, no annotations.
uses a 12-dimensional vocoder trained with a subset of JVS, fine tuned from a multivoice model.
this model uses a neural spline flow likelihood.