dasLLAMA text-to-speech models

The speech models dasLLAMA serves, converted to GGUF, and the 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, and kokoro-ru, the Russian fine-tune of Kokoro-82M - behind one facade and one das-native text front end (an English normalizer, a part-of-speech tagger and a grapheme-to-phoneme pass; a Russian normalizer, a stress pass over the stress pack and a letter-to-sound 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 packs.

Files

file what bytes sha256
kitten-nano.gguf KittenTTS nano 0.8, f32 59331456 4556948c36a29e4be5ad521e597a20ae817059c404fea3c5d935afa73506d9da
kitten-nano-q8.gguf KittenTTS nano 0.8, the fast form: the weights the Q8_0 lane serves, stored as that lane mints them - no mint at load 16646016 ed1b412ccc10499d2f4bc4fe4f4c78c7308dfb772c311782989f41ed8aee9dbe
kitten-nano-kq.gguf KittenTTS nano 0.8, the small form: the recipe lane's weights stored the same way - Q4_K and Q6_K convs beside the Q8_0 groups 14785920 d92014bdfa7cc7bafb3093f3e755c0fdc76a974e03c363b1b6c42b610860326e
kitten-mini.gguf KittenTTS mini 0.8, f32 295975008 e127a95c6ffa390eccc5956e7156ff055e9f30078cf16361cbcd0eea6c21fdb5
kitten-mini-q8.gguf KittenTTS mini 0.8, the fast form: the weights the Q8_0 lane serves, stored as that lane mints them - no mint at load 80495296 1e1d094ee91aaf28a33ec4a1617a1acdd7d3a35fba7a681d2d99e022532701ed
kitten-mini-kq.gguf KittenTTS mini 0.8, the small form: the recipe lane's weights stored the same way - Q4_K and Q6_K convs beside the Q8_0 groups 59725504 f90e4f5aaac96cbd887fce4b9eeb3d05bd39a34a9f80c188a32a8902e6475d46
kokoro-82m.gguf Kokoro-82M v1.0 with its 54 voice packs, f32 352965024 e0d6584b5b650730bd62e7b105a2e9502384bbe3fbebf3030113560355e6232f
kokoro-82m-q8.gguf Kokoro-82M v1.0 with its 54 voice packs, the fast form: the weights the Q8_0 lane serves, stored as that lane mints them - no mint at load 95792544 cb1ade7573957ed00bffe4968f1cc0ddec4222dd28d654636b93b5ea406a6d20
kokoro-82m-kq.gguf Kokoro-82M v1.0 with its 54 voice packs, the small form: the recipe lane's weights stored the same way - Q4_K and Q6_K convs beside the Q8_0 groups 71391648 4b5102ab80b568ed6d9dc23417d7864fb23435b6f744729146a8fd1977b5bc4f
kokoro-ru.gguf kokoro-ru v2, the Russian fine-tune of Kokoro-82M, the base checkpoint with its two voice packs sveta and masha, f32 325804992 635a43debfcbff5cf1fd0380a4ffc55dfacff083ae16d057bae6d7e7c9a29c9f
kokoro-ru-q8.gguf kokoro-ru v2 with sveta and masha, the fast form: the weights the Q8_0 lane serves, stored as that lane mints them - no mint at load 88575552 1961e4ca0cfde51c6e57244e19c45163c3c7e67808d94387cee942849789e5ff
kokoro-ru-kq.gguf kokoro-ru v2 with sveta and masha, the small form: the recipe lane's weights stored the same way - Q4_K and Q6_K convs beside the Q8_0 groups 64174656 b3a238b377514b11fa778ba7b8003f83c460f3a427a5fea54525c2064d725bdb
kokoro-ru-dima.gguf kokoro-ru v2, the second checkpoint with its one male voice pack dima, f32 325282688 7720e1087c0c9720577e7e0b383dd31f75b7f6ab45bcd88df1f28df3c08c56a6
kokoro-ru-dima-q8.gguf kokoro-ru v2 with dima, the fast form: the weights the Q8_0 lane serves, stored as that lane mints them - no mint at load 88436768 17ae7fb6698d66011facffc7e9a3e4b7db84d17542d4b9523967417545fa5d2a
kokoro-ru-dima-kq.gguf kokoro-ru v2 with dima, the small form: the recipe lane's weights stored the same way - Q4_K and Q6_K convs beside the Q8_0 groups 64035872 b396a6c2dfe4461b7127eb89aade5007654f39606e31cb56a31fa489870d978c
tts_g2p_ru.bin the Russian stress pack: every inflected form of the OpenRussian dictionary, the forms and readings the English and Russian Wiktionary add and the names Russian Wikipedia marks, each with its stressed vowel (a form with two readings carrying both, the more frequent one first), the ё spellings a text written without ё needs, and the rules that pick a reading from the words beside it - a stress (з+амок / зам+ок) or a spelling (все / всё) 15387767 74b8f3c4d3a3e140b7c7937d15da95a37328e7e3347392088514e47e68e5265d
ru-stress-rubert-base-q4.gguf the Russian stress-context model: reads a sentence and picks the reading of each word whose stress (зам+ок / з+амок) or ё spelling (все / всё) depends on the context; ruBert-base fine-tuned on a private dataset, Q4_K with a Q8_0 embedding table. Optional: copy it beside a Russian voice as tts_ctx_ru.gguf 120522080 28692fe9fbea9cb7b06f6c39e18a78e4c1e30b37173f2040808865a1d7730624
ru-stress-ruelectra-small-q4.gguf the same model in the small form, ruElectra-small fine-tuned on a private dataset; used the same way 38428800 43475b1535fdcd3d142bb04ac845909f2557bee32b524af5e19eb1528bbc1ae3
tts_g2p.bin the grapheme-to-phoneme pack: misaki's gold and silver lexicons in both English dialects (one merged table per tier), CMUdict 0.7a rendered into the American inventory, the g2p_en GRU spelling model 14011554 d7b6afea7a0901a877d10531054d3f967845c58f415f0ea937decd8a3081ccd5
tts_g2p_en_us.bin the same pack with the American tier alone - no British values, CMUdict pruned of every word the American lexicon carries; the web serving set's pack, British voices are not offered on it 10257455 6f69d2e74565bd7d876b8d1f4042bf8c1c5b615387fa26ff45215cf447932154
tts_postag.bin the tokenizer exception table and the averaged-perceptron PTB tagger 12566510 38c2e85f7fef3e57d561d2aa0af25fccda4276376ba1993c3dbc2ae0ebfa57b4
pocket-tts-en-q8.gguf Kyutai Pocket TTS, English (english_2026-04): the served GEMM weights as Q8_0, the rest f16, its unigram tokenizer and 19 voices as latent frames inside; reads text, needs no pack 136812032 ff41c88d0099375e096c1a0bb9db43c10e8b53b8bc9d3874b03f658f94a5487e
pocket-tts-de-q8.gguf Pocket TTS German (6 layers), the same form, one voice (juergen) 136519648 1762eb997ab113247a5083d55dfd22bf5ee83d8256016fce13a67ba4f3fb6fd9
pocket-tts-es-q8.gguf Pocket TTS Spanish (6 layers), one voice (lola) 136518688 65ab208f0ecc976e752889bd3a1c1f61b51a95515e1159fde8468177ba083128
pocket-tts-it-q8.gguf Pocket TTS Italian (6 layers), one voice (giovanni) 136515072 65507c2ff6c87836bca5ea50e66e55a12ecec78390556efae75eb35b4c44f754
pocket-tts-pt-q8.gguf Pocket TTS Portuguese (6 layers), one voice (rafael) 136521024 80ae1a859658149843de6701fc4b86868c34925b55490c21429c497d333201b6
pocket-tts-fr-q8.gguf Pocket TTS French (24 layers, the only French model Kyutai ships), one voice (estelle) 377328544 edb510a0c441f50817b8aed8cfec961b77bd9ee12691c9db9f91427a966cb738
pocket-tts-de-kq.gguf Pocket TTS German in the small form (the backbone and the codec transformers Q4_K, the flow head Q8_0), one voice (juergen) 77466176 29d3088f3db5aa9958b589e60da023b6c89a80a341986eb43cbe7aa29bbd7432
pocket-tts-es-kq.gguf Pocket TTS Spanish in the small form (the backbone and the codec transformers Q4_K, the flow head Q8_0), one voice (lola) 77465216 23b3a52cc77d4efaf291f46507a5e0412896f87e1f71fd0d4afa3c4d2e9d0008
pocket-tts-it-kq.gguf Pocket TTS Italian in the small form (the backbone and the codec transformers Q4_K, the flow head Q8_0), one voice (giovanni) 77461600 dfe62eccdf774c660a8c11dc7da1778a3e1f986c245f4d37b3bedd4c7f21c309
pocket-tts-pt-kq.gguf Pocket TTS Portuguese in the small form (the backbone and the codec transformers Q4_K, the flow head Q8_0), one voice (rafael) 77467552 0c25e851662d09fb782f1eac11caf38e7b00f1f19f307577700c0b8c2014decd
pocket-tts-fr-kq.gguf Pocket TTS French in the small form (the backbone and the codec transformers Q4_K, the flow head Q8_0), one voice (estelle) 205028864 f56ac84a47687d517a0a4e1776532096503567d09250b090c50c9a74d0c55cb8
pocket-tts-en-kq.gguf Pocket TTS English in the small form: the backbone and the codec transformers as Q4_K, the flow head as Q8_0, the codec convolutions f16, the embedding table Q4_K; its tokenizer, the codec encoder (so it clones) and the 19 voices as latent frames 77758560 9b549549e44601435ec5df3f48b5ef977cf1449bcf92e6ae49db4b07da20e757
pocket-tts-en-stuart-kq.gguf the same small form with one voice (stuart_bell) as latent frames and no codec encoder: reads text in that voice, cannot clone 67005056 e9e69e52254f48e544509a0dff3b594779f737eac5b372d61adda8b696a5af46
pocket-tts-en-18m-q8.gguf Kyutai Pocket TTS English 18.2M (english_18m_2026-10: an 8-layer 512-wide backbone, a 128-wide decoder behind the 100M model's encoder), the served GEMMs Q8_0, the rest f16, the 19 voices and the encoder inside 36712480 d5b138be971d7907eb7a568b827510d85775d7ce933fad4dba0245d2d0536fa2
pocket-tts-en-18m-kq.gguf the 18.2M model in the small form: the backbone and the text table Q4_K, the flow head and the decoder transformer Q8_0 23547232 ed8db0741c377ea21a3603101d9215b538e2c143b1916f459e1b599ce096716b
pocket-tts-en-18m-stuart-kq.gguf the 18.2M small form with one voice (stuart_bell) as latent frames and no codec encoder: reads text in that voice, cannot clone 12793728 3678b27db407fa0715feb7c0863c63c1cb00a28f9387fa76f8e4922e30c8f124
pocket-tts-en-7m-q8.gguf Pocket TTS English 6.9M (english_7m_2026-10: a 224-wide grouped-query backbone), the served GEMMs Q8_0 22991840 b3d99bc412613c4323f023572c782b980a375fed83d88239b7e5d68de2621c4c
pocket-tts-en-2m.gguf Pocket TTS English 1.9M (english_2m_2026-10: a 112-wide backbone, a 4-layer decoder), every tensor f16 - its widths sit off the quant lattices 23758400 612a583ce4811f8904200295fa9293423074cffe4eadbcc3a5232aec1be5f68d
pocket-tts-en-18m-refl-q8.gguf the 18.2M model after Kyutai's reward fine-tuning (english_18m_refl_2026-10), the q8 form 36712480 3460c9cff8102e997369ab3ee7e32e1d020a8f1cffc4a454dc1ecb059f51d4dc
pocket-tts-en-7m-refl-q8.gguf the 6.9M model after reward fine-tuning (english_7m_refl_2026-10), the q8 form 22991840 80ca90b4a60716693722ed19a91945d326f4611f8573defe27f722d6f4bd50a2
pocket-tts-en-2m-refl.gguf the 1.9M model after reward fine-tuning (english_2m_refl_2026-10), f16 23758432 54dead7cb04aa87971b32f4a25288d566c95c739290581df91742155fcc61230

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 -jit utils/dasllama-server/cli.das -- speak --tts kitten-nano.gguf --voice expr-voice-2-f -o hello.wav "Hello."

Pocket TTS English is the cloning model: 137 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 3.36 / UTMOS 4.395 on an Apple M5 Max (measured 2026-11-16 with the module's harness/tts_rig.py, the engine under the JIT tier on the Metal seats, the kernels untuned - DASLLAMA_ALLOW_UNTUNED=1). 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, and each has its small form beside it (pocket-tts-<code>-kq.gguf: 77 MB for the six-layer models, 205 MB for French). 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, 78 MB: the backbone and the codec transformers as Q4_K, the flow head as Q8_0, the codec convolutions f16, the embedding table Q4_K, the encoder and the 19 voices inside (the rig row at alba: WER 3.45 / UTMOS 4.384, the same box, rig and date); 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, 67 MB (WER 3.36 / UTMOS 4.106 at that voice): it reads text in that voice and cannot clone. The 18.2M model's twins of these two files - pocket-tts-en-18m-kq.gguf (24 MB) and pocket-tts-en-18m-stuart-kq.gguf (13 MB) - are the same forms at half the bytes and twice the browser speed; the pages keep the 100M model, whose pace reads more naturally.

The 2026-10 small English models are Kyutai's distilled Pocket TTS at 18.2M, 6.9M and 1.9M parameters (english_18m_2026-10, english_7m_2026-10, english_2m_2026-10; the -refl files are the same models after Kyutai's reward fine-tuning), each with the same 19 voices and the 100M model's codec encoder inside, so each clones. Every Pocket file here is served the same way: the voice clip is trimmed to end on a pause, the end-of-speech logit is ignored on a chunk's first six frames, each chunk fades in over 5 ms, and a trailing comma or dash becomes a period. On the 200-sentence rig at alba on an Apple M5 Max (measured 2026-10-23, harness/tts_rig.py, the JIT tier, the CPU chain - the GPU seats decline these shapes): pocket-tts-en-18m-q8.gguf reads WER 3.09 / UTMOS 4.341, pocket-tts-en-18m-refl-q8.gguf 3.04 / 4.369, pocket-tts-en-18m-kq.gguf 3.18 / 4.232, pocket-tts-en-7m-q8.gguf 3.41 / 4.357, pocket-tts-en-7m-refl-q8.gguf 3.86 / 4.377, pocket-tts-en-2m.gguf 5.32 / 3.777 and pocket-tts-en-2m-refl.gguf 4.68 / 3.823; the reference package on the same sentences reads 5.41 / 4.364 (18m), 5.59 / 4.378 (7m) and 9.18 / 3.817 (2m) - the WER gap is the English normalizer in front of our files (the reference reads "1995" as digits), the UTMOS gap the q8 lane's. On that box, served through dasllama-server on the Metal seats (one sentence, eight requests), the 100M q8 file answers in 50 ms for 4.6 s of speech (a real-time factor of 0.011), the 18m files in 28-30 ms for 4.4-4.6 s (0.006), the 7m in 28 ms for 5.0 s (0.006) and the 2m in 25 ms for 4.7 s (0.005) - every stage of every file on the GPU seats. The reference package on the same box and sentence, one CPU thread as it pins, takes 325 ms on the 100M and 156 ms on the 18m. Every figure of this paragraph is the Pocket small-model leg's entry of PERF_LEDGER.md, with its instruments and commands. In the browser (the parrot page's wasm64 build under Chrome on the same box, seven workers) pocket-tts-en-18m-kq.gguf runs at 0.079 against the 100M kq file's 0.161 - twice as fast, twelve times faster than real time - and clones a 6.5 s voice in 0.47 s against 0.87.

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. kokoro-ru carries the Russian voices - sveta and masha in kokoro-ru.gguf, dima in kokoro-ru-dima.gguf - and needs tts_g2p_ru.bin beside it; with the English packs beside it too, a Latin word inside Russian text is read in English. Russian stress comes from the pack, or from the text itself: + before the stressed vowel (зам+ок) or U+0301 after it, which always wins, so a homograph the pack picks the other way can be fixed in the text. With tts_ctx_ru.gguf beside the voice - a copy of either ru-stress-*-q4.gguf - a word with two readings is stressed by its sentence; without it the pack's rules and its more frequent reading decide. A text written without ё gets the letter back: always where the е spelling is no word (еще, объем), and where both spellings are words (все / всё) by the sentence with tts_ctx_ru.gguf, by the pack's rules without it. dasllama-server --tts <directory holding the packs> serves the front end with no voice: POST /v1/audio/phonemes answers the text with + before each stressed vowel and its phonemes, for a synthesis that runs elsewhere. 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.

Each of the five phoneme-family files also ships in two forms that load with no mint - the engine reads the quantized weights as the file stores them. <name>-q8.gguf is the fast form: the weights the default Q8_0 lane serves. <name>-kq.gguf is the small form: the decoder, the generator and the text encoder's convolutions as Q4_K and the upsamplers as Q6_K beside the Q8_0 groups. Kokoro-82M is 96 MB and 71 MB against the f32 file's 353; kokoro-ru 89 and 64 against 326; kokoro-ru-dima 88 and 64 against 325; Kitten mini 80 and 60 against 296; Kitten nano 17 and 15 against 59. On the 200-sentence rigs (measured 2026-11-16 on an Apple M5 Max with harness/tts_rig.py, the JIT tier, the engine's CPU chain - DASLLAMA_METAL_TOWER=0 - with the kernels untuned (DASLLAMA_ALLOW_UNTUNED=1), the Russian set under --lang ru) the files read, fast then small, WER % / UTMOS: kokoro-ru sveta 0.70 / 3.835 and 0.70 / 3.838, masha 0.65 / 3.715 and 0.65 / 3.714, kokoro-ru-dima dima 0.70 / 3.739 and 0.74 / 3.743, Kokoro-82M af_heart 2.82 / 4.500 on both, Kitten nano Bella 3.13 / 3.986 and 3.41 / 3.956, Kitten mini Bella 2.86 / 4.341 and 3.13 / 4.323. The f32 files on the same lanes, minted at load (the same rig and box the same day, the kernels tuned), read 0.74 / 3.844 and 0.65 / 3.846 (sveta), 0.65 / 3.716 on both (masha), 0.74 / 3.740 and 0.78 / 3.743 (dima), 2.68 / 4.500 on both (Kokoro-82M), 3.41 / 3.966 and 2.91 / 3.941 (Kitten nano), 2.86 / 4.339 and 3.00 / 4.325 (Kitten mini): one word of either set is 0.04 points, and no form leads. The f32 files stay the reference - they serve the same lanes after a mint at first load, and take a lane pin, which a file that stores its lane does not.

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).
  • kokoro-ru.gguf / kokoro-ru-dima.gguf: converted by modules/dasLLAMA/harness/convert_kokoro.py (--checkpoint, --config, --voices, --lang ru) from zaakirio/kokoro-ru at d649c57b239b18c4c384378127cbf01dba039bc1 (kokoro-ru-v2-base.pth with voices/sveta.bin and voices/masha.bin; kokoro-ru-v2-dima.pth with voices/dima.bin).
  • <name>-q8.gguf / <name>-kq.gguf of the five files above: written from the f32 file of the same name by utils/dasllama-convert --model <name>.gguf --gguf <name>-q8.gguf (and --recipe for -kq) - the same weights, each quantized group stored as the engine's Q8_0 or recipe lane mints it.
  • tts_g2p_ru.bin: modules/dasLLAMA/harness/build_g2p_ru_data.py over the OpenRussian dictionary data (Badestrand/russian-dictionary: nouns.csv, adjectives.csv, verbs.csv, others.csv), the forms and readings of the English and Russian Wiktionary (the machine-readable extracts of kaikki.org) and the stress Russian Wikipedia's lead sentences mark, the pronoun inflections the dictionary leaves out added by the builder, the homograph defaults ordered by a frequency table.
  • ru-stress-rubert-base-q4.gguf / ru-stress-ruelectra-small-q4.gguf: converted by modules/dasLLAMA/harness/convert_ru_stress.py from ai-forever/ruBert-base and ai-forever/ruElectra-small, each fine-tuned on a private dataset; the fine-tuned checkpoints are not distributed.
  • 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 under pocket-tts 4.0.1 from the files its english_2026-04 config pins: kyutai/pocket-tts languages/english_2026-04/model.safetensors at 19f95fe2df36e79fbd9f10008595cc4c977a0fcc and kyutai/pocket-tts-without-voice-cloning languages/english_2026-04/tokenizer.json at 00eac05ed3d16bdc3f6b5d598874019c34a89214; the voice clips from kyutai/tts-voices at 323332d33f997de8394f24a193e1a76df720e01a (voice-zero/, voice-donations/, vctk/, alba-mackenna/casual.wav), each trimmed to end on a pause and stored as the codec encoder's latent frames.
  • pocket-tts-en-kq.gguf / pocket-tts-en-stuart-kq.gguf, and the 18.2M model's pocket-tts-en-18m-kq.gguf / pocket-tts-en-18m-stuart-kq.gguf: the same sources through modules/dasLLAMA/harness/convert_pocket.py --kq (the one-voice files 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.
  • pocket-tts-<code>-q8.gguf and pocket-tts-<code>-kq.gguf for German, Spanish, Italian, Portuguese and French: the same converter (--q8, --kq) under pocket-tts 4.0.1 from the files each language's config pins (german, spanish, italian, portuguese, french_24l): kyutai/pocket-tts languages/<config>/model.safetensors at 3e82814a68665eec246ff649b14c71331f955c06 and kyutai/pocket-tts-without-voice-cloning languages/<config>/tokenizer.json at 4e1e0a3e611c51c0b4ed8174fc10f32a54644303 (French at 8843db76457a91db32077edf8dfcd1c0e3e755fd); the language's default clip from kyutai/pocket-tts (estelle from kyutai/tts-voices), stored the same way.
  • pocket-tts-en-18m-q8.gguf, pocket-tts-en-7m-q8.gguf, pocket-tts-en-18m-refl-q8.gguf, pocket-tts-en-7m-refl-q8.gguf (--q8), pocket-tts-en-2m.gguf, pocket-tts-en-2m-refl.gguf (f16) and pocket-tts-en-18m-kq.gguf (--kq): converted by modules/dasLLAMA/harness/convert_pocket.py under pocket-tts 4.0.1 from kyutai/pocket-tts languages/<config>/model.safetensors at db04f0e653473e5c7333c69686ac92b7460b0c1d (the configs english_18m_2026-10, english_7m_2026-10, english_2m_2026-10 and their _refl_ twins) and each config's tokenizer.model of kyutai/pocket-tts-without-voice-cloning at 08ffc58775e62aad9409511b9591c9af626270f3; the same 19 clips, each trimmed to end on a pause as the package's prompt path trims it, then through the codec encoder (the 100M model's, which these bundles carry unchanged).

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

file licence terms
kitten-nano.gguf, kitten-mini.gguf and their -q8 / -kq forms Apache-2.0 KittenML's weights, converted; LICENSE.APACHE-2.0; the architecture is StyleTTS2 (MIT, LICENSE.STYLETTS2)
kokoro-82m.gguf and its -q8 / -kq forms Apache-2.0 hexgrad's weights and voices, converted; the architecture is StyleTTS2 (MIT, LICENSE.STYLETTS2)
kokoro-ru.gguf, kokoro-ru-dima.gguf and their -q8 / -kq forms OpenRAIL the kokoro-ru weights and voices, converted - the source repository tags them openrail and carries no licence text of its own; the voices come from the Dialogs corpus, which its card calls OpenRAIL; the architecture is StyleTTS2 (MIT, LICENSE.STYLETTS2)
tts_g2p_ru.bin CC BY-SA 4.0 the OpenRussian dictionary data (openrussian.org), the English and Russian Wiktionary and Russian Wikipedia (LICENSE.UD_EWT carries the licence text); tts_g2p_ru.bin.LICENSE names each source
ru-stress-rubert-base-q4.gguf Apache-2.0 fine-tuned from ai-forever's ruBert-base, converted (LICENSE.APACHE-2.0)
ru-stress-ruelectra-small-q4.gguf MIT fine-tuned from ai-forever's ruElectra-small, converted; MIT as its model card states
tts_g2p.bin Apache-2.0 and BSD-2-Clause misaki and g2p_en (Apache-2.0), CMUdict (LICENSE.CMUDICT, Carnegie Mellon University)
tts_postag.bin CC BY-SA 4.0 the tagger weights are trained on UD English-EWT (LICENSE.UD_EWT); the exception table and the silver tags come from spaCy (MIT, LICENSE.SPACY); Gutenberg prose is public domain
pocket-tts-*.gguf (every Pocket file, the -kq ones included) CC BY 4.0 Kyutai's weights and tokenizer, converted (LICENSE.CC-BY-4.0); the reference implementation is MIT (LICENSE.POCKET_TTS) and not included; the voice clips: voice-zero and voice-donations CC0, VCTK (CSTR, University of Edinburgh) and Alba Mackenna CC BY 4.0 - the sidecar lists each

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.

Downloads last month
2,224
GGUF
Model size
76.6M params
Architecture
kitten-tts
Hardware compatibility
Log In to add your hardware

We're not able to determine the quantization variants.

Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support