Aava TTS · LoRA Adapter

Persian speech synthesis adapter for Orpheus 3B, trained from nearly 2,500 hours of Persian source audio.

آوا یک مدل تبدیل متن فارسی به گفتار است که بر پایهٔ Orpheus 3B تنظیم شده است.

This repository contains the PEFT/LoRA weights. For a self-contained Transformers model, use Aava 3B.

Model card


Direct base	`unsloth/orpheus-3b-0.1-ft-unsloth-bnb-4bit`
Method	LoRA, rank 128, alpha 256, dropout 0.05
Target modules	attention, MLP projections, embeddings
Context	4,096 tokens
Audio codec	SNAC 24 kHz
Language	Persian (`fa`)
Adapter storage	~1.77 GB

Load

import torch
from peft import PeftModel
from unsloth import FastLanguageModel

model, tokenizer = FastLanguageModel.from_pretrained(
    model_name="unsloth/orpheus-3b-0.1-ft-unsloth-bnb-4bit",
    max_seq_length=4096,
    dtype=torch.bfloat16,
    load_in_4bit=True,
)
model = PeftModel.from_pretrained(model, "KEYHAN-A/aava-tts-persian")
model.eval()

Generation produces Orpheus audio tokens that must be redistributed and decoded with SNAC. The complete reference implementation is in Aava Refinement.

Training

Setting	Value
Training records	492,596
Tokenized corpus	~1,905.3 hours
Epochs / steps	3 / 184,725
Effective batch	8
Learning rate	1.5e-4
Scheduler	cosine with restarts
Optimizer	AdamW 8-bit
Precision	bfloat16
Hardware	RTX 3090 24 GB
Training time	~11 days
Reported eval loss	4.2045

The often-cited ~2,400-hour number is the raw acquisition duration, not the duration of the refined token corpus.

Training corpus

Training used 547,328 multi-source, multi-speaker Persian records under one conditioning label, aava.

The reported evaluation loss is an internal training metric. A post-training audit found exact transcript duplication across the original sequential splits, so it must not be interpreted as a clean generalization benchmark.

Intended use and limitations

Persian TTS research, prototyping, and accessibility experiments.
Persian-only training; other languages are unsupported.
Output stability and voice similarity vary with prompts and sampling.
The conditioning label does not identify one specific speaker.
Do not use the model for impersonation, fraud, deceptive media, harassment, or identity abuse.
No formal MOS, speaker-similarity, memorization, or production-latency benchmark is published for v1.0.0.

License

The adapter weights are released under Apache-2.0, following the direct Orpheus/Unsloth base. Orpheus is Apache-2.0 and SNAC is MIT licensed. No endorsement by the upstream projects is implied.

Citation

@misc{aava_tts_2026,
  title   = {Aava TTS: Persian Orpheus 3B Adapter},
  author  = {{KEYHAN-A}},
  year    = {2026},
  version = {1.0.0},
  url     = {https://huggingface.co/KEYHAN-A/aava-tts-persian}
}