Dataset · amaai-lab
Studio speech with disfluencies and emotions for TTS
amaai-lab/DisfluencySpeech
Nearly 10 hours of studio English speech from a single speaker with paralinguistics for training natural speech synthesis.
What's inside
A set of several thousand audio-text pairs where a single speaker re-recorded in studio quality about 10 hours of live informal utterances from the Switchboard-1 telephone corpus. The recordings contain natural pauses, hesitations, and paralinguistic elements.
What you can do with it
- Training text-to-speech (TTS) models capable of predicting and voicing live pauses and emotions directly from text
- Fine-tuning ASR systems for better recognition of spontaneous speech with disfluencies
- Size
- 1K–10K rows
- Downloads this month
- 608
- Likes
- 25
Models trained on it
- mradermacher/LvW-Dictate-1.5B-GGUF 416
- flockspeech/goldfinchText → speech 80
- Lucasgitguy-co/LvW-Dictate-1.5BText generation 26
- nazarkozak/whisper-small-disfluent-verbatim-loraAutomatic speech recognition 17
- nazarkozak/whisper-small-disfluent-smoothed-loraAutomatic speech recognition 12
- pradachan/whisper-large-v3-turbo-disfluency-loraAutomatic speech recognition 12
- amaai-lab/DisfluencySpeech_BenchmarkBText → speech 8
- amaai-lab/DisfluencySpeech_BenchmarkCText → speech 6
How to load it
from datasets import load_dataset
ds = load_dataset("amaai-lab/DisfluencySpeech")