openpecha/garchen_rinpoche_benchmark_result

Tibetan Speech Recognition Model Performance Report

Overview

This report summarizes the performance evaluation of a Wav2Vec2-based speech recognition model trained on Tibetan speech data of Garchen Rinpoche. The evaluation was conducted across different training checkpoints using multiple error metrics.

Training datasets

train

Benchmark Dataset Information

• Total audio duration: 1.07 hours
• Number of audio segments: 893
• Average segment duration: 4.31 seconds
• Content type: Teaching material
• Speaker age group: 70-90 years

Model Architecture

• Base Model: ganga4364/mms_300_v4.96000
• Model Type: Wav2Vec2ForCTC
• Fine-tuning Method: Parameter-Efficient Fine-Tuning (PEFT) with LoRA

Training Parameters

• Batch Size: 8 (per device)
• Gradient Accumulation Steps: 2
• Learning Rate: 3e-4
• Training Epochs: 100
• Warmup Steps: 500
• FP16 Training: Enabled
• Evaluation Strategy: Steps-based
  • Logging Interval: 100 steps
  • Evaluation Interval: 1000 steps
  • Save Interval: 1000 steps
• Save Total Limit: 50 checkpoints
• Data Loading: 4 workers
• Monitoring: Weights & Biases (wandb)

Training Progress

Character Error Rate (CER) across training checkpoints:

Checkpoint	Micro CER (%)
Base model	27.67
5000 steps	27.41
10000 steps	23.37
19000 steps	22.93

Final Model Performance (Checkpoint 19000)

Word Error Rate (WER)

• Micro-average WER: 39.42%
• Macro-average WER: 45.92%

Error Analysis

• Total test sentences: 893
• Error breakdown:
  • Substitutions: 4,217
  • Insertions: 779
  • Deletions: 1,190

Sample Prediction

Example transcription:

Reference:
དེའི་རྩིས་གཞི་ད་ད་ལྟ་སྐད་ཆ་བཤད་དགོས་རེད་ད། དྲན་པས་ཡང་མི་གནོད།

Model Prediction:
དེ་རིང་དེ་དུ་དག་ད་ྟ་སྐད་ཆ་བཤད་དགོས་རེད། ད་དྲིརིང་གི་ཡང་མི་འདུག

CER for this example: 39.34%

Key Findings

1. Consistent Improvement: The model shows steady improvement in CER across training checkpoints, with a total reduction of 4.74 percentage points from base model to final checkpoint.

2. Character vs Word Accuracy: While the final character-level error rate is relatively low (19.46%), the word-level error rate is higher (39.42%), indicating challenges in maintaining word integrity during recognition.

3. Error Distribution: The majority of errors are substitutions (4,217), followed by deletions (1,190) and insertions (779), suggesting the model is more prone to replacing characters than inserting or deleting them.

Conclusion

The model demonstrates promising performance for Tibetan speech recognition, particularly at the character level. However, the higher word error rate suggests room for improvement in capturing complete word structures. Future work might focus on reducing the gap between character and word-level accuracy.

---

Report generated on July 14, 2025

---

Dataset Statistics

Configuration: default

Split: train

Total Rows: 893

audio_duration

• Type: numerical
• Data Type: float64
• Sum: 3,870.19
• Average: 4.33

strata

• Type: categorical
• Data Type: object
• Unique Values: 8

Value Distribution:

Value	Count	Percentage
70-80__medium__Prayer	125	14.00%
70-80__short__Teaching	123	13.77%
70-80__long__Teaching	123	13.77%
70-80__long__Prayer	121	13.55%
70-80__medium__Teaching	118	13.21%
70-80__long__Q&A	108	12.09%
80-90__long__Practice	108	12.09%
70-80__medium__Q&A	67	7.50%

age_group

• Type: categorical
• Data Type: object
• Unique Values: 2

Value Distribution:

Value	Count	Percentage
70-80	785	87.91%
80-90	108	12.09%

duration_category

• Type: categorical
• Data Type: object
• Unique Values: 3

Value Distribution:

Value	Count	Percentage
long	460	51.51%
medium	310	34.71%
short	123	13.77%

content_type

• Type: categorical
• Data Type: object
• Unique Values: 4

Value Distribution:

Value	Count	Percentage
Teaching	364	40.76%
Prayer	246	27.55%
Q&A	175	19.60%
Practice	108	12.09%

base_cer

• Type: numerical
• Data Type: float64
• Sum: 247.09
• Average: 0.28

cer_5000

• Type: numerical
• Data Type: float64
• Sum: 244.76
• Average: 0.27

cer_10000

• Type: numerical
• Data Type: float64
• Sum: 208.71
• Average: 0.23

cer_19000

• Type: numerical
• Data Type: float64
• Sum: 204.79
• Average: 0.23

---