ta4tsering's picture
Update README with dataset statistics (#2)
e9ed5e6 verified
metadata
dataset_info:
  features:
    - name: file_name
      dtype: string
    - name: url
      dtype: string
    - name: Dilgo_kyentse_Rinpoche_model_inference
      dtype: string
    - name: audio_duration
      dtype: float64
    - name: Human_transcript
      dtype: string
    - name: General_model_inference
      dtype: string
    - name: situ_rinpoche_model_inference
      dtype: string
  splits:
    - name: train
      num_bytes: 791538
      num_examples: 552
  download_size: 266432
  dataset_size: 791538
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*

Tibetan ASR Model Comparison: Garchen Rinpoche Dataset

Dataset Information

  • Content: Garchen Rinpoche Tibetan teachings
  • Samples: 552 audio segments
  • Duration: 1.21 hours total
  • Metrics used: Character Error Rate (CER), Word Error Rate (WER), Syllable Error Rate (SER)

Model Performance Summary

General STT Model

  • Model: OpenPecha General Tibetan STT
  • Character Error Rate: 27.53%
  • Word Error Rate: 57.24%
  • Syllable Error Rate: 51.15%
  • Major Error Type: Word substitutions (75%)

Situ Rinpoche Model

  • Model: OpenPecha Situ Rinpoche ASR
  • Character Error Rate: 28.92%
  • Word Error Rate: 60.68%
  • Syllable Error Rate: 53.68%
  • Major Error Type: Word substitutions (75%)

Dilgo Khyentse Model

  • Model: OpenPecha Dilgo Khyentse ASR
  • Character Error Rate: 65.96%
  • Word Error Rate: 92.63%
  • Syllable Error Rate: 91.55%
  • Major Error Type: High proportion of deletions (33%)

Detailed Documentation here


Dataset Statistics

Configuration: default

Split: train

Total Rows: 552

audio_duration

  • Type: numerical
  • Data Type: float64
  • Sum: 4,362.82
  • Average: 7.90