CoolFace
Datasetpublic

chosenek/czech-speech-combined

Czech Speech Combined Dataset Quality-filtered Czech speech dataset for TTS/ASR training. 146,153 clips across ~1,700 speakers from 5 sources. Sources Source Clips Hours Speakers Origin audiobooks 24,535 ~33h 13 Czech audiobook narrations audiobooks_new 28,131 ~39h 8+ Czech audiobook narrations yodas_czech 28,208 ~37h ~2,300 YODAS YouTube speech (quality-filtered) voxpopuli_czech 12,679 ~33h 45 VoxPopuli parliament speech commonvoice_czech 52… See the full description on the dataset page: https://huggingface.co/datasets/chosenek/czech-speech-combined.

sourceHugging Facecc-by-4.0updated 4mo agoView on Hugging Face
2likes170downloads
Dataset Card

Czech Speech Combined Dataset

Quality-filtered Czech speech dataset for TTS/ASR training. 146,153 clips across ~1,700 speakers from 5 sources.

Sources

SourceClipsHoursSpeakersOrigin
audiobooks24,535~33h13Czech audiobook narrations
audiobooks_new28,131~39h8+Czech audiobook narrations
yodas_czech28,208~37h~2,300YODAS YouTube speech (quality-filtered)
voxpopuli_czech12,679~33h45VoxPopuli parliament speech
commonvoice_czech52,600~67hmanyMozilla Common Voice

Quality Pipeline

All audio passed through:

  • —Duration filter (2-20s)
  • —SNR analysis (min 10)
  • —Silence floor check
  • —Voice activity ratio
  • —Pacing (chars/sec) check
  • —Edge cut detection
  • —Sentence completeness filter (terminal punctuation)
  • —Whisper large-v3 transcription + confidence filter (YODAS, VoxPopuli)

Audio Format

  • —24 kHz mono WAV, 16-bit PCM (embedded in Parquet)
  • —Columns: audio (struct with bytes+path), text, source, filename

Usage

python
from datasets import load_dataset

ds = load_dataset("chosenek/czech-speech-combined", split="train")
print(ds[0])  # {'audio': {...}, 'text': '...', 'source': '...', 'filename': '...'}

# Filter by source
audiobooks = ds.filter(lambda x: x["source"] == "audiobooks")