Aava TTS · 3B Merged
Ready-to-load Persian Orpheus model with the Aava LoRA merged into the base, trained from nearly 2,500 hours of Persian source audio.
آوا یک مدل تبدیل متن فارسی به گفتار است که خروجی ۲۴ کیلوهرتز تولید میکند.
Quick start
pip install torch transformers snac soundfile numpy
curl -L https://raw.githubusercontent.com/KEYHAN-A/aava-refinement/main/examples/inference.py -o aava_inference.py
python aava_inference.py --text "سلام، من آوا هستم"
The script auto-selects CUDA, Apple MPS, or CPU and downloads
KEYHAN-A/aava-tts-persian-3b on first use.
Model card
| Architecture | LlamaForCausalLM + SNAC decoder |
| Parameters | ~3.78B |
| Precision | bfloat16 |
| Storage | ~7.59 GB, two safetensors shards |
| Context | 4,096 tokens |
| Codec / output | SNAC, mono 24 kHz |
| Language | Persian (fa) |
This model is the inference-oriented counterpart of the Aava LoRA adapter.
Training
| Setting | Value |
|---|---|
| Training records | 492,596 |
| Tokenized corpus | ~1,905.3 hours |
| Method | LoRA, rank 128, alpha 256 |
| Epochs / steps | 3 / 184,725 |
| Effective batch | 8 |
| Learning rate | 1.5e-4 |
| Precision | bfloat16 |
| Reported eval loss | 4.2045 |
The ~2,400-hour raw-source total must not be confused with the refined training duration.
Training corpus
The training corpus contains 547,328 multi-source, multi-speaker Persian
records under one conditioning label (aava).
The reported eval loss is an internal training metric. Exact transcript duplicates cross the original sequential splits, so the metric is not a clean generalization benchmark.
Intended use and limitations
- Persian TTS research, prototyping, and accessibility experiments.
- Other languages and code-switching are unsupported.
- Long generations can repeat or fail to terminate.
- Output voice identity and similarity have not been formally evaluated.
- No v1.0.0 MOS, memorization, latency, or speaker-similarity benchmark exists.
- Do not use for impersonation, fraud, deceptive media, harassment, or identity abuse.
License
The merged weights are Apache-2.0, following the direct Orpheus/Unsloth base. Orpheus is Apache-2.0 and SNAC is MIT licensed. No upstream endorsement is implied.
Related
Citation
@misc{aava_tts_merged_2026,
title = {Aava TTS: Persian Orpheus 3B Merged Model},
author = {{KEYHAN-A}},
year = {2026},
version = {1.0.0},
url = {https://huggingface.co/KEYHAN-A/aava-tts-persian-3b}
}