treadon's picture
Upload README.md with huggingface_hub
4121821 verified
metadata
language:
  - en
license: cc-by-4.0
task_categories:
  - text-to-speech
tags:
  - dac
  - audio-tokens
  - speech
  - tts
  - codebook
  - descript-audio-codec
  - librispeech
  - 16khz
pretty_name: Speech DAC Tokens 16kHz (2 Codebooks)
size_categories:
  - 10K<n<100K

Speech DAC Tokens 16kHz (2 Codebooks)

Pre-tokenized speech dataset using DAC at 16kHz with 2 codebooks. Optimized for speech TTS training — 16kHz captures the full speech frequency range without wasting capacity on inaudible frequencies.

Why 16kHz?

  • Speech lives below 8kHz — 16kHz sample rate is sufficient (Nyquist)
  • 50 tokens/sec per codebook vs 87 at 44kHz — shorter sequences, faster training
  • 2 codebooks at 16kHz produce intelligible speech — verified by listening tests
  • No resampling needed — LibriSpeech is natively 16kHz

Dataset Summary

Stat Value
Total samples 132,479
Total audio ~464 hours
Source LibriSpeech clean-100 + clean-360
Language English
DAC model 16kHz, 2 of 12 codebooks
Codebook size 1,024 entries each
Tokens per second 100 (50/codebook x 2)
Token sequence length 149-2,047 (mean: 1,327)

Format

Column Type Description
text string Original text transcription
prompt string {text}<|audio_start|><|c1_X|><|c2_Y|>...<|audio_end|>
input_ids list[int] Pre-tokenized with Qwen3-0.6B + 2cb DAC tokens
attention_mask list[int] All 1s
labels list[int] Copy of input_ids
n_audio_frames int Number of DAC time frames
n_tokens int Total token count

Audio tokens interleaved: c1, c2, c1, c2, ... per frame.

Related

Citation

@inproceedings{panayotov2015librispeech,
  title={Librispeech: an ASR corpus based on public domain audio books},
  author={Panayotov, Vassil and Chen, Guoguo and Povey, Daniel and Khudanpur, Sanjeev},
  booktitle={ICASSP},
  year={2015}
}