weights-and-wires/driven-data

On Top of Pasketti — Children's Word ASR (Training Data)

Training dataset for the DrivenData "On Top of Pasketti" Children's Word ASR competition.

Contents

A single train.parquet file containing 95,572 rows with the following columns:

Column	Type	Description
utterance_id	string	Unique identifier for each utterance
child_id	string	Anonymized speaker identifier
session_id	string	Recording session identifier
audio_path	string	Original relative path to the .flac file
audio_duration_sec	float64	Duration of the audio clip in seconds
age_bucket	string	Age range: 3-4, 5-7, 8-11, 12+, or unknown
md5_hash	string	MD5 checksum of the original audio file
filesize_bytes	int64	Size of the original audio file in bytes
orthographic_text	string	Normalized orthographic transcription
audio_bytes	binary	Raw FLAC audio bytes

Quick Start

import pyarrow.parquet as pq

# Metadata only (fast)
df = pq.read_table("train.parquet", columns=["utterance_id", "orthographic_text", "age_bucket"]).to_pandas()

# Stream with audio in batches
pf = pq.ParquetFile("train.parquet")
for batch in pf.iter_batches(batch_size=256, columns=["utterance_id", "audio_bytes", "orthographic_text"]):
    audio_data = batch.column("audio_bytes")
    transcripts = batch.column("orthographic_text")
    # process ...

Statistics

Stat	Value
Total utterances	95,572
Total audio duration	~185.4 hours
Compression	zstd
File size	~11.9 GB

Age Distribution

Age Bucket	Count	%
3–4	10,112	10.6%
5–7	11,490	12.0%
8–11	73,970	77.4%