Gooya RizehPizeh v1.5

Provenance

• Original model: Piper VITS, warm-started from the Mana Persian Piper checkpoint (epoch=6012-step=4203520.ckpt, sdp enabled). This model is therefore a fine-tune of Piper, continuing from a mature Persian training run rather than training from scratch.
• Front end: phonemic input produced by Negara v7.1 G2P (grapheme-to-phoneme), with phoneme ids mapped through the Mana 256-symbol inventory (157 real phonemes).
• Training data: AvaSanj clean-core v2 — 102,584 utterances whose phoneme labels were rebuilt by the OOF (out-of-fold) listener policy:
  • oof_listener_winner … 48,656 (OOF AvaSanj ASR margin ≥ 0.1)
  • stored_audio_prompt … 42,244 (unchanged approved prompts)
  • three_listener_consensus … 11,342 (unanimous multi-listener rows)
  • human_override … 41
  • human_reviewed_v71_overlay … 301
  • 28,253 rows changed vs. the stored prompt (the G2P improvement delivered by this project).
• Split: 5% validation, num_test_examples: 0.

Model

• Generator parameters: 23,663,792 (~23.7 M)
• Architecture (Piper/VITS): hidden_channels 192, filter_channels 768, inter_channels 192, 6 flow layers, 2 attention heads, resblock 2, upsampling rates [8, 8, 4] (upsample initial channel 256), mel_channels 80, use_sdp true, num_symbols 256, num_speakers 1.
• Vocab: 157 phoneme tokens (text phoneme type, Mana id map with ^/_/$ control tokens).

Inference

Files

• gooya-fa.onnx — ONNX model (inference runtime)
• gooya-fa.onnx.json — Piper voice/config metadata
• checkpoint/epoch=*-val_mel=*.ckpt — PyTorch training checkpoint (resumable)