Modelpublic

prasadvittaldev/orpheus-telugu-male-qlora-v1-GGUF

sourceHugging Faceotherupdated 2mo agoView on Hugging Face

0likes60downloads

Model Card

Orpheus Telugu Male (QLoRA v1) -- GGUF

A Telugu text-to-speech voice, finetuned from Orpheus. Voice prefix: `telugu_male`.

Eval loss 3.1779 -- the best figure recorded in this project, below the Tamil full-finetune baseline of 3.235 that had needed a rented A100. This one trained on a single consumer 16 GB card at no GPU cost.

Measured median F0 of generated speech: 136.9 Hz.

See also the matching female voice and the GGUF build.

Model details


Base	canopylabs/3b-hi-pretrain-research_release (Orpheus, Llama-3.2-3B backbone)
Method	QLoRA (4-bit nf4 base, bf16 compute), r=64 alpha=128, attention + MLP; embeddings frozen
Schedule	2 epochs, lr 2e-4 cosine, batch 1 x grad-accum 16
Hardware	one 16 GB RTX 5060 Ti, trained locally
Corpus	SYSPIN Telugu (arpit-tiwari/syspin-telugu-tts) -- studio read speech with ground-truth transcripts

How it works

Orpheus emits SNAC audio-codec tokens which a SNAC vocoder decodes to 24 kHz mono PCM. Prompt format is {voice}: {text}; the voice is a literal text prefix learned at training time, so this model has no zero-shot cloning and takes no audio input.

Audio token layout: control tokens 128257-128262, audio offset 128266, 7 tokens per frame each offset by (position_in_frame 0-6) * 4096.

Inference

Recommended sampling: temperature=0.3, repetition_penalty=1.3. Output is quiet (peaks ~0.2-0.4) -- peak-normalise or level it downstream.

Licence and provenance

Derived from the SYSPIN Telugu corpus. The mirror this was trained from does not declare a licence; anyone intending commercial use should confirm terms with the upstream corpus holders (IISc SPIRE Lab / Bhashini) rather than relying on this repo.

Files

Quantisation	Size	Notes
f16	6.2 GB	conversion source
Q8_0	3.3 GB	served in production
Q4KM	2.0 GB	measured faster than Q8_0 (RTF 0.90 vs 1.09) at no audible cost
Q2_K	1.4 GB	included for comparison only

Runs with llama.cpp / llama-cpp-python. You still need a SNAC vocoder (hubertsiuzdak/snac_24khz) to turn the emitted codec tokens into audio -- the GGUF alone produces tokens, not waveforms.