Datasetpublic

dlyog/af_heart_arm_tts_dataset

af_heart_arm_tts_dataset The distillation corpus used to train dlyog/af_heart_arm_tts — a single-voice TTS model built for Arm inference on NVIDIA DGX Spark. 3.0000 hours · 4,104 clips · 24 kHz mono 16-bit PCM, synthesized by Kokoro-82M speaking af_heart. This is knowledge distillation: the teacher generated every clip, so the student can approach it but never exceed it. The pipeline that produced this — and that reproduces it for any other voice — is at… See the full description on the dataset page: https://huggingface.co/datasets/dlyog/af_heart_arm_tts_dataset.

sourceHugging Faceapache-2.0updated 2mo agoView on Hugging Face

0likes183downloads

Dataset Card

afheartarmttsdataset

The distillation corpus used to train [dlyog/af_heart_arm_tts](https://huggingface.co/dlyog/af_heart_arm_tts) — a single-voice TTS model built for Arm inference on NVIDIA DGX Spark.

3.0000 hours · 4,104 clips · 24 kHz mono 16-bit PCM, synthesized by Kokoro-82M speaking af_heart. This is knowledge distillation: the teacher generated every clip, so the student can approach it but never exceed it.

The pipeline that produced this — and that reproduces it for any other voice — is at [github.com/dlyog/arm-cooperative-tts](https://github.com/dlyog/arm-cooperative-tts)

Layout

metadata.csv        <wav>|<text>, one clip per line
wavs/000001.wav     24 kHz mono 16-bit
wavs/000002.wav
...

metadata.csv is pipe-separated — the format the trainer reads directly:

000001.wav|The quick brown fox jumps over the lazy dog.
000002.wav|Pack my box with five dozen liquor jugs.

Fetch

bash

git clone https://github.com/dlyog/arm-cooperative-tts
cd arm-cooperative-tts
bash 1_SyntheticAudioDataset/manage.sh fetch

Or directly:

python

from huggingface_hub import snapshot_download
snapshot_download("dlyog/af_heart_arm_tts_dataset", repo_type="dataset",
                  local_dir="1_SyntheticAudioDataset/output")

How the text was chosen

source	sentences	why
word bank	62	pangrams, spoken numbers, contractions, punctuation — deliberately the cases a phoneme-level model gets wrong. Every pangram is verified programmatically to contain all 26 letters, not trusted from memory
CMU ARCTIC	1,132	phonetically balanced prompts selected by CMU's Language Technologies Institute from out-of-copyright Project Gutenberg prose, specifically for single-speaker TTS
LLM-generated	remainder	topical variety — small talk, directions, scheduling, customer service — so the voice is not only read-aloud prose

Deduplicated case-insensitively.

How the audio was made

Kokoro-82M af_heart, run in-process, at 24 kHz to match the student model's sample rate.

Clips longer than 6.25 s were dropped, not truncated. A cut-off sentence teaches a model to stop mid-phrase. 6.25 s is therefore the longest clip here, and the trained model is not validated beyond that length in a single pass — longer text should be chunked by sentence.

Statistics


clips	4,104
total duration	3.0000 h
sample rate	24,000 Hz
channels	mono
bit depth	16-bit PCM
longest clip	6.25 s
voice	`af_heart` (Kokoro-82M)
language	English (US)

Building your own

Nothing downstream is tied to af_heart. One argument changes the speaker:

bash

bash 1_SyntheticAudioDataset/manage.sh all 3.0 am_michael

The text corpus needs no network and no external service — the word bank and CMU ARCTIC prompts are offline. The LLM expansion step is optional and is skipped automatically when no endpoint is configured.

Licensing and provenance

Apache-2.0. The audio was generated by Kokoro-82M (Apache-2.0), so the clips carry no copyleft obligation. CMU ARCTIC prompts derive from out-of-copyright Project Gutenberg text, distributed by CMU for research use.

This dataset is synthetic. It contains no recordings of any real person; the speaker is a synthetic voice from Kokoro-82M. This work is not affiliated with or endorsed by the Kokoro-82M authors.

Intended use: distilling a small single-voice TTS model. Because the teacher is a hard quality ceiling, this corpus is not suitable for training a model intended to surpass Kokoro-82M, nor for speaker-verification or voice-cloning research involving real identities.

Citation

bibtex

@misc{chawdhury2026afheartdataset,
  author       = {Chawdhury, Tarun Kumar},
  title        = {af\_heart\_arm\_tts\_dataset: A 3-hour Synthetic
                  Single-Voice TTS Distillation Corpus},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/datasets/dlyog/af_heart_arm_tts_dataset}},
  organization = {DLYog Lab Research Services LLC}
}