Fabio
c36510c21e
2.4 KiB
XTTS v2
Coqui XTTS v2 provides offline multilingual text-to-speech with a speaker reference.
| Field | Value |
|---|---|
| Family | xtts_v2 |
| Task / mode | tts / offline |
| Languages | en, es, fr, de, it, pt, pl, tr, ru, nl, cs, ar, zh-cn, hu, ko, ja, hi |
| Voice input | Required reference WAV; 6–30 seconds recommended |
| Output | Mono 24 kHz audio |
Weights
Packages in the model spec download from
onicubag/xTTSv2-GGUF.
Choose a package containing one GPT GGUF and xtts-v2-vocoder-f16.gguf. Q8_0 is the default; the original GPT file and Q4_K_M are also available.
mel_norms.bin is required only when neither GGUF embeds normalization
metadata or a mel_stats tensor.
Pass the directory containing these files, or its GPT GGUF, as --model.
Usage
audiocpp_cli --task tts --family xtts_v2 --model models/XTTS-v2-Q8 \
--backend cuda --threads 8 \
--text "Hello, my name is Anna and this is a test." \
--language en --voice-ref reference.wav \
--out out.wav --log
Request options: language, temperature, top_k, top_p,
repetition_penalty, max_mel_tokens, seed, text_chunk_size, and
text_chunk_mode. Set them using
--request-option name=value. Defaults and limits are in the model spec.
The reference is mixed to mono and resampled internally. Repeated requests
within a session reuse prepared conditioning. Set
--session-option xtts_v2.conditionals_cache_slots=2 to retain two voices;
the default is 1, and 0 disables caching. The least recently used voice is evicted
when the slots are full.
Use --text-chunk-size N and --text-chunk-mode default|tag_aware|japanese|endline
to control the framework text chunker. Defaults are 256 Unicode codepoints and
word-budget splitting. Chunking modes do not add model support for emotion or
other text tags.
Validated scope
The local validation covers English offline CUDA, Vulkan and CPU generation, repeated requests, seed reproducibility, long text, changed references, stereo reference mixdown, and German UTF-8 tokenization. It does not establish full multilingual text normalization parity: upstream's Chinese/Japanese/Korean romanization and non-English number normalization are not implemented here.
License
The model weights use the Coqui Public Model License.