dasllama-tts

dasLLAMA text-to-speech models

The speech models dasLLAMA serves, converted to GGUF, and the two front-end packs the phoneme families load. dasLLAMA is the daslang inference engine (https://github.com/GaijinEntertainment/daScript, modules/dasLLAMA); the models are the StyleTTS2 lineage - KittenTTS nano and mini, Kokoro-82M - behind one facade and one das-native text front end (a normalizer, a part-of-speech tagger, a grapheme-to-phoneme pass), and Kyutai's Pocket TTS, a continuous-audio language model that reads text and clones a voice from a clip. No espeak-ng, no phonemizer: the front end is data, and the data is in the two packs.

Files

The packs sit beside whichever GGUF you load; the loader reads them from the model's directory - tts_g2p.bin when it is there, else tts_g2p_en_us.bin. The GGUFs carry f32 weights: dasLLAMA quantizes the served layouts to Q8_0 at first load and keeps the result beside the file as a prepared image, so the f32 file is also the reference lane.

Use

dasllama-server --tts kitten-nano.gguf          # POST /v1/audio/speech, the OpenAI shape
daslang utils/dasllama-server/txt2wav.das -- --tts kitten-nano.gguf --voice expr-voice-2-f --text "Hello." --out hello.wav

Pocket TTS English is the cloning model: 152 MB, 19 voices (alba the default, bill_boerst, caro_davy, anna, george, ...), and any voice from a few seconds of 24 kHz audio through tts_register_voice. It reads text, so it needs neither pack; the English normalizer runs in front of it. On the 200-sentence rig at alba this file reads WER 3.91 / UTMOS 4.328 at a real-time factor of 0.051 on an Apple M1 Max, against the reference package's 5.00 / 4.393 / 0.210 (measured 2026-09-09 with the module's harness/tts_rig.py, the engine under the JIT tier with the box's tune profile, the reference package under torch on one thread). The five other languages are the same form, one file each with Kyutai's default clip for that language as its only voice (German juergen, Spanish lola, Italian giovanni, Portuguese rafael, French estelle); the German, Spanish, Italian and Portuguese files are the six-layer models, French exists only as the 24-layer one. A voice cloned from any clip speaks the file's language with the clip's accent. Text in those languages is read as it is, since the normalizer is English. pocket-tts-en-kq.gguf is the English model in the small form, 75 MB: the backbone and the codec transformers as Q4_K, the flow head and the codec convolutions as Q8_0, the embedding table Q4_K, the encoder and the 19 voices inside (on the rig at alba: WER 3.86 / UTMOS 4.295 at a real-time factor of 0.049 on the same box); it is the file the browser examples on dasllama.io fetch. pocket-tts-en-stuart-kq.gguf is that form with one voice, stuart_bell, and no codec encoder, 65 MB: it reads text in that voice and cannot clone.

Kitten nano is the phoneme families' served default: 59 MB, eight voices, a real-time factor of 0.03 on an Apple M1 Max (measured 2026-09-02 with the same rig). Its voices are expr-voice-2-m through expr-voice-5-f, and the upstream's alias names (Bella, Jasper, ...) are accepted for them. Kokoro ships 54 packs, of which the front end drives the 28 English ones: the 20 American (af_*, am_* - af_heart, am_adam, ...) and the 8 British (bf_*, bm_* - bf_emma, bm_george, ...), each phonemized in its own dialect. The other 26 packs speak languages the front end does not phonemize yet, so they are not listed and are refused. Some Kokoro packs, the British ones most audibly, render a comma or a full stop as a breath; the reference pipeline does the same, so it is the pack, not the engine. Every model speaks at 24 kHz. GET /v1/stats on the server lists the served model's voices.

Provenance

• kitten-nano.gguf / kitten-mini.gguf: converted by modules/dasLLAMA/harness/convert_kitten.py from KittenML/kitten-tts-nano-0.8 at 7a1db645b1f3ab9420761d87428e042b9cec3f26 and KittenML/kitten-tts-mini-0.8 at c02725660cea441db4c383af69f1f26f5cd00947 (the ONNX weights, the voice table, the config).
• kokoro-82m.gguf: converted by modules/dasLLAMA/harness/convert_kokoro.py from hexgrad/Kokoro-82M at f3ff3571791e39611d31c381e3a41a3af07b4987 (kokoro-v1_0.pth and the voice packs).
• tts_g2p.bin: modules/dasLLAMA/harness/build_g2p_data.py over misaki 0.9.4 (us_gold.json, us_silver.json, gb_gold.json, gb_silver.json), CMUdict 0.7a (the NLTK distribution) and g2p_en 2.1.0 (checkpoint20.npz), plus harness/g2p_local_additions.json.
• tts_postag.bin: modules/dasLLAMA/harness/train_postag.py over UD English-EWT and Project Gutenberg prose tagged by spaCy's en_core_web_sm; the tokenizer exception table is spaCy's.
• pocket-tts-en-q8.gguf: converted by modules/dasLLAMA/harness/convert_pocket.py --q8 from kyutai/pocket-tts languages/english_2026-04/model.safetensors at 19f95fe2df36e79fbd9f10008595cc4c977a0fcc and the tokenizer of kyutai/pocket-tts-without-voice-cloning at d29db7978e464fb90cb3359ee0c69a273b9142cc; the voice clips from kyutai/tts-voices at 323332d33f997de8394f24a193e1a76df720e01a (voice-zero/, voice-donations/, vctk/, alba-mackenna/casual.wav).
• pocket-tts-en-kq.gguf / pocket-tts-en-stuart-kq.gguf: the same sources through modules/dasLLAMA/harness/convert_pocket.py --kq (the second with --voices stuart_bell --no-cloning); the K-quant blocks are ggml's own quantizer, and each voice is stored as the latent frames of its clip through the model's codec encoder.

The whole set is rebuilt by modules/dasLLAMA/performance/build_tts_data.das. Parity against the reference implementations (block by block, and the front end sentence by sentence on a 200-sentence corpus) is the test suite under modules/dasLLAMA/tests/test_tts_*.das.

Licences, per file

Each .LICENSE sidecar beside a file names its sources; the full texts are in this repository. The engine that reads these files is under the daslang licence in its own repository.