Datasets:

TalhaAhmed
/

gal-fe01-clean

language:
  - ur
license: mit
task_categories:
  - text-to-speech
  - automatic-speech-recognition
tags:
  - urdu
  - tts
  - audio
  - cleaned
pretty_name: TalhaAhmed/gal-fe01-clean — Cleaned Urdu TTS
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
      - split: validation
        path: data/validation-*
dataset_info:
  features:
    - name: audio
      dtype:
        audio:
          sampling_rate: 22050
    - name: transcript
      dtype: string
  splits:
    - name: train
      num_bytes: 8551969899
      num_examples: 26693
    - name: validation
      num_bytes: 452930073
      num_examples: 1406
  download_size: 7656933617
  dataset_size: 9004899972

TalhaAhmed/gal-fe01-clean

Cleaned version of TalhaAhmed/gal-fe01 Generated on 2026-03-24 11:14 UTC by clean_urdu_tts_dataset.py.

Cleaning summary

Issue type	Rows affected	Action
Word / phrase repetition	1,868 (+ 1,547 phrase)	First occurrence kept
Character-level stutter	9	Collapsed to 2 chars
Noise / filler tokens	194	Removed
Stray ASCII words	0	Stripped
Garbage / non-Urdu rows	0	Dropped
Too short after cleaning	717	Dropped

Row counts

	Count
Original rows	28,846
Rows kept	28,099
Rows dropped	747 (2.6%)
Rows with text fixed (kept)	1,397

Schema

Identical to the source dataset. Only the text (transcript) column is modified. The audio column and all other columns are untouched.

A full per-row audit trail is available in cleaning_report.csv in this repo.

Source

<TalhaAhmed/gal-fe01> — TalhaAhmed/gal-fe01