--- license: cc-by-4.0 language: [en] pretty_name: kale-tts-1-data task_categories: [text-to-speech] size_categories: [10K text`. It runs on vLLM 0.33.0 in bf16 on Modal, 4 H100 GPUs in parallel (one per container, 8 CPU cores and 64 GiB each), 10,000 prompts per call. vLLM batches all of a call's prompts on the GPU at once: about 14x faster than plain generation, 255x real time per H100, about $0.02 of GPU time per hour of audio. Loading the model takes 80-110 s, so big calls are what keep it cheap. 46,000 takes (118 hours of audio) took about 17 minutes on 4 H100s. Sampling: temperature 0.4, top-p 0.9, repetition penalty 1.1, one seed per take. SNAC 24 kHz ([hubertsiuzdak/snac_24khz](https://huggingface.co/hubertsiuzdak/snac_24khz) @ `d73ad176`) turns Maya-1's codes into audio. 3. **Filter.** Every take is measured and kept only if it passes every check (next sections). The measuring ran on up to 8 GPUs (H100 or L40S), the rough-ending measure on up to 64 CPU containers. 4. **Books.** Only takes from books that are clearly public domain in the EU and the US go in. 5. **Split.** Sentences that appear together in any take form one group; a group goes to `test` by a hash of its text, so no `test` sentence is spoken anywhere in `train`, in any voice. The audio is Maya-1's output as it came out: no denoising, trimming or loudness change. The scripts are in the `recipe/render/` and `recipe/filter/` folders of [kale-tts-1-tiny](https://huggingface.co/KaleComputer/kale-tts-1-tiny). ## What's in it ![What's in kale-tts-1-data](diagrams/datacontents.png) - `data//train-NNNNN-of-MMMMM.parquet` and `data//test-00000-of-00001.parquet`: 55 files, each up to about 0.5 GB, 19.9 GB in all. - `preview/preview-00000-of-00001.parquet`: the 200 preview takes. - `metadata/metadata.parquet` (12.6 MB): every column except `audio` and `maya_codes`, for filtering without downloading audio. - `metadata/books.json`: the source books with the death years that decide their status. `metadata/voices.json`: the voice prompts. | column | meaning | |---|---| | `audio` | 24 kHz mono FLAC (16-bit), converted from Maya-1 + SNAC's float32 output; every take peaks at or below 0.975, so nothing clips | | `maya_codes` | the take's raw SNAC codes as Maya-1 produced them (frames x 7, uint16, token id minus 128266); SNAC 24 kHz decodes them back to the original float audio | | `id`, `voice`, `split` | take id, voice id (the config name above), split | | `text` | what was read | | `kind`, `n_sentences` | `single` (one sentence) or `passage` (2-5 consecutive sentences of one paragraph), and how many sentences | | `description` | the voice prompt the take was rendered from | | `book_id`, `book_title`, `book_author` | the source book | | `seconds`, `wpm`, `f0_median`, `voiced_share`, `peak` | length, spoken words a minute, median pitch in Hz, voiced share, peak level | | `lead_silence_s`, `tail_silence_s`, `end_level_db` | silence before and after the speech, level of the last frames (a cut-off take ends loud) | | `whisper_text`, `whisper_missed_words`, `whisper_extra_words` | Whisper large-v3-turbo's read-back of the take | | `end_token` | Maya-1 ended the take with its own end token (filled in where the render logged it) | | `voice_similarity` | ECAPA-TDNN (`speechbrain/spkrec-ecapa-voxceleb`) cosine of the take to its voice's centre, on the released audio; higher = closer | | `render_round` | the render round the take came from | | `in_release_training` | the take trained kale-tts-1-nano and kale-tts-1-tiny | ## How we selected the takes A take is kept only if all of these hold: | check | rule | |---|---| | complete | Maya-1's own end token | | read-back | Whisper large-v3-turbo hears exactly the text: no missed and no extra words | | edges | 20 ms to 0.8 s of silence before, at most 0.8 s after; not cut off (last frames at or below -15 dB); no full-scale sample | | pace | 100-230 spoken words a minute (Kenzie: up to 270, her natural pace) | | pitch | median pitch within 15 % of the voice's median | | voice match | a speaker embedding close to the voice's own centre (the mean of its closest 60 % of takes) | | sentence ends | no rough (creaky or cut-off) ending: under 100 ms of rough frames in the last ~300 ms of every sentence, inside passages too; sentence ends found by forced alignment; pitch floor 110 Hz for women, 70 Hz for men. One rough inner end drops the whole passage. | Most takes don't make it. In the render rounds behind this data, the filter kept 115,246 of 284,122 rendered takes (41 %). A take can fail more than one check; the most common: | check failed | takes | |---|---| | read-back: missed words | 71,077 | | read-back: extra words | 70,846 | | pace | 64,676 | | rough ending | 58,012 | | pitch | 17,987 | | end edge | 5,185 | | voice match | 1,368 | | start edge | 1,217 | | no end token | 581 | | cut off | 313 | Men need more renders: their takes fail the sentence-end check more often. After the filter, the book rule leaves out 18,551 more kept takes (40.59 h) from six translated books whose translators' public-domain status we can't show. Rejected takes are not in this dataset. ## Use it ```python from datasets import load_dataset ds = load_dataset("KaleComputer/kale-tts-1-data") # all 10 voices, train + test jack = load_dataset("KaleComputer/kale-tts-1-data", "d13", split="train", streaming=True) # one voice, streamed for take in jack.take(3): print(take["text"], take["seconds"], take["wpm"]) ``` Filter first without downloading audio: ```python import pandas as pd meta = pd.read_parquet("hf://datasets/KaleComputer/kale-tts-1-data/metadata/metadata.parquet") fast = meta[(meta.voice == "s057b") & (meta.wpm <= 230)] ``` Train on `train`, evaluate on `test`: its sentences never appear in `train`. About 25 hours of clean takes per voice is where a small model's clarity levels off; 5 hours already sounds right but is less clear. Train on whole passages and speak sentence by sentence: passages carry the melody. `maya_codes` lets you train codec-token models on the exact Maya-1 tokens without re-encoding. What it trained: | model | parameters | CPU file (int8) | WebGPU file (float32) | Galaxy Tab S10 Lite | unclear words vs Maya-1 | |---|---|---|---|---|---| | [kale-tts-1-base](https://huggingface.co/KaleComputer/kale-tts-1-base) | 81.8M | 327 MB (PyTorch) | - | not built for the tablet | +3.9 points | | [kale-tts-1-tiny](https://huggingface.co/KaleComputer/kale-tts-1-tiny) | 45.7M | 79.9 MB | 183.2 MB | 24x real time | +1.4 points | | [kale-tts-1-nano](https://huggingface.co/KaleComputer/kale-tts-1-nano) | 7.5M | 14.8 MB | 30.6 MB | 44x real time | +1.6 points | kale-tts-1-base is our teacher: Kokoro-82M fine-tuned on these 10 voices. tiny keeps its planner, nano is distilled from it, and both learn their sound straight from this audio. ## Limits - English only, US and British accents. Book narration, mostly older prose: no dialogue, no numbers, little modern vocabulary. The voices read text; they don't converse. - One voice prompt per voice: no speaker variety inside a voice. - The filter is automatic. A small share of takes may still carry a mispronunciation Whisper heard as right, or a rough ending; the deep male voices keep some creak at sentence ends. - Kenzie has 17.25 hours here; the models trained on 5.41 of them. Her rows go up to 270 words a minute, her trained rows stay within 230. - The models also trained on 17,282 takes from the six held-back books; those takes aren't in this dataset. - The accent words in the voice prompts are Maya-1's own vocabulary. The voices are what Maya-1 made of them, not recordings of any community, and may not sound like real speakers of those accents. ## Licence and credits Audio, codes and metadata: CC BY 4.0. Texts: public domain in the EU and the US: the deciding person (author, or translator for a translation) died in 1955 or earlier, and the book was published before 1931. AI-generated speech: every take is synthetic speech made by an AI model, in voices that belong to no real person. Label anything made from this data as AI-generated, never present it as a human recording, and never use it to imitate a real person. Credits: based on Maya-1: all audio is original speech from Maya-1 (Maya Research, Apache-2.0), decoded with SNAC (Hubert Siuzdak, MIT). Used for scoring only, nothing of them is in the data: Whisper large-v3-turbo (OpenAI, MIT) for the read-back, ECAPA-TDNN `speechbrain/spkrec-ecapa-voxceleb` (SpeechBrain, Apache-2.0) for `voice_similarity`. Acknowledgements: the source books were made available by Project Gutenberg (www.gutenberg.org).