Quantised GGUF builds of prasadvittaldev/orpheus-kannada-female-qlora-v1, a Kannada single-speaker TTS voice for Orpheus.

The LLM emits SNAC audio-codec tokens; a SNAC vocoder decodes them to 24 kHz mono PCM. You need both this model and the SNAC decoder — this file alone does not produce audio.

Q2_K is deliberately not published. Measured on sibling checkpoints, 2-bit K-quant was slower than Q4_K_M despite being 30% smaller (dequantisation overhead exceeds the bandwidth saved) and inflated output length by up to 45%, indicating unstable stopping. It costs quality for negative speed.

Prompt format is kannada_female: <text>, wrapped in Orpheus control tokens (start 128259, tail [128009, 128260, 128261, 128257]). Audio tokens start at offset 128266, 7 tokens per SNAC frame, each + (position_in_frame 0-6) * 4096.

Recommended sampling: temperature 0.3, repetition_penalty 1.3. Output is quiet (the source corpus is low-level studio audio) — peak-normalise at inference.

QLoRA (4-bit nf4 base, r=64 / alpha=128 on attention + MLP, embeddings frozen) over canopylabs/3b-hi-pretrain-research_release. The Hindi speech pretrain is used as the base because its SNAC audio-token embeddings are already trained — essential here, since embeddings stay frozen.

Speech data originates from the SYSPIN project (IISc SPIRE Lab); please refer to the upstream corpus for its terms. Base model by Canopy Labs. The Apache-2.0 tag above covers this repository's own artifacts, not the underlying corpus.

Single speaker, single language. Trained on read studio speech, so highly expressive or conversational delivery is out of distribution. Numerals and English loanwords in Kannada script are handled but benefit from text normalisation upstream.

SYSPIN Kannada ships no speaker ID. Single-speaker was established by fingerprinting (gender, age, experience) across shards spanning the corpus: every female clip is the same speaker (39, 16 years), so filtering on gender yields one voice rather than a blend. Generated output measures 210.5 Hz median F0 against the male model's 136.4 Hz, with no overlap.

This speaker reads more slowly than the male (10.8 vs 13.5 characters/second) and her source recordings are quieter (peak 0.15 vs 0.56) -- normalise or level at inference.