rleo/XTTS-v2-GGUF

XTTS-v2 (French, English) — GGUF for xtts-rs

Single-file GGUF conversion of coqui/XTTS-v2 for xtts-rs, a Rust/Candle port of Coqui XTTS-v2 (no Python, no PyTorch). The file holds the config, the BPE tokenizer, the 58 built-in voices and the weights: GPT in q4k, HiFi-GAN decoder in F32 (weight norm folded).

License: Coqui Public Model License 1.0.0 — non-commercial use only. These weights are a conversion of Coqui's XTTS-v2 and stay under the CPML; its full text is in LICENSE.txt (also at https://coqui.ai/cpml). Anyone you give a copy of the model or of its output must get these terms too.

Not for llama.cpp. xtts-rs layout (general.architecture = xtts).

file	size	note
xtts-v2-q4k.gguf	276 MB	--gpt-dtype q4k --no-cloning: the voice-cloning encoders are left out (cloning is not implemented in xtts-rs)

Other dtypes (q5k, q6k, q8_0, f16) and the cloning encoders can be written from the original checkpoint with xtts convert (below); they bring no measurable quality gain.

Built-in voices (Coqui's speakers_xtts.pth): Claribel Dervla, Ana Florence, Daisy Studious, Gracie Wise, Tammie Ema, Alison Dietlinde, Annmarie Nele, Asya Anara, Brenda Stern, Gitta Nikolina, Henriette Usha, Sofia Hellen, Tammy Grit, Tanja Adelina, Vjollca Johnnie, Andrew Chipper, Badr Odhiambo, Dionisio Schuyler, Royston Min, Viktor Eka, Abrahan Mack, Adde Michal, Baldur Sanjin, Craig Gutsy, Damien Black, Gilberto Mathias, Ilkin Urbano, Kazuhiko Atallah, Ludvig Milivoj, Suad Qasim, Torcull Diarmuid, Viktor Menelaos, Zacharie Aimilios, Nova Hogarth, Maja Ruoho, Uta Obando, Lidiya Szekeres, Chandra MacFarland, Szofi Granger, Camilla Holmström, Lilya Stainthorpe, Zofija Kendrick, Narelle Moon, Barbora MacLean, Alexandra Hisakawa, Alma María, Rosemary Okafor, Ige Behringer, Filip Traverse, Damjan Chapman, Wulf Carlevaro, Aaron Dreschner, Kumar Dahl, Eugenio Mataracı, Ferran Simen, Xavier Hayasaka, Luis Moray, Marcos Rudaski.

Quality and speed

Greedy audio codes from the F32 checkpoint are identical to coqui-tts 0.27.5; the decoder matches at 74-79 dB SNR. Word error rate (Parakeet TDT 0.6B v3, 2 voices × 2-3 seeds):

GGUF	French WER	English WER
q4k, xtts-rs (final period dropped)	0.6 %	0.0 %
q8_0 (with the final period, as in Coqui)	3.1 %	0.2 %

RTX 4090 (CUDA 12): 776 MiB VRAM idle, 1032 MiB peak (one ~20 s utterance), ~13x real time, first streamed audio 30-45 ms. CPU: ~0.4x real time.

French and English are validated. The other XTTS languages (es, de, it, pt, pl, tr, ru, nl, cs, ar, hu, hi) run through the same code but were not checked; zh-cn, ja and ko are not supported by xtts-rs.

Usage

hf download rleo/XTTS-v2-GGUF xtts-v2-q4k.gguf --local-dir models

# xtts-rs (https://github.com/rzafiamy/xtts-rs, build with ./build.sh --cuda)
xtts speak -m models/xtts-v2-q4k.gguf -l fr -v "Ana Florence" "Bonjour à tous." -o bonjour.wav
xtts serve -m models/xtts-v2-q4k.gguf --port 8091 --lang fr   # /v1/audio/speech, /stream

# or convert the original checkpoint yourself
xtts convert ./XTTS-v2 -o models/xtts-v2-q4k.gguf --gpt-dtype q4k --no-cloning

zallama serves it as the xtts-server backend (/v1/audio/speech, /v1/realtime).

Credits

XTTS-v2 by Coqui (model, weights, voices), under the Coqui Public Model License 1.0.0. Conversion and Rust port: rzafiamy/xtts-rs.