Update README.md

README.md

@@ -9,9 +9,6 @@ tags:
 - wav2vec2
 - wav2vec2-xls-r
 - burmese
-- myanmar
-- ctc
-- hf-asr-leaderboard
 datasets:
 - custom
 metrics:
@@ -19,70 +16,97 @@ metrics:
 - wer
 - chrf
 pipeline_tag: automatic-speech-recognition
+model-index:
+- name: Wav2Vec2-XLS-R-300M-ASR
+  results:
+  - task:
+      type: automatic-speech-recognition
+      name: Automatic Speech Recognition
+    metrics:
+    - name: Validation CER
+      type: cer
+      value: 3.53
+    - name: Validation WER
+      type: wer
+      value: 24.25
+    - name: Validation chrF
+      type: chrf
+      value: 94.97
+    - name: Test CER (Unseen Speakers)
+      type: cer
+      value: 6.58
+    - name: Test WER (Unseen Speakers)
+      type: wer
+      value: 33.24
+    - name: Test chrF (Unseen Speakers)
+      type: chrf
+      value: 87.69
 ---
 
-# Wav2Vec2 XLS-R (300M) — Burmese Speech Recognition
+# Wav2Vec2-XLS-R-300M-ASR ( Burmese )
 
-Fine-tuned **facebook/wav2vec2-xls-r-300m** (~317M parameters) with a Connectionist Temporal Classification (CTC) head on a curated 22.38-hour Burmese speech corpus.
+This model is a fine-tuned version of [facebook/wav2vec2-xls-r-300m](https://huggingface.co/facebook/wav2vec2-xls-r-300m) for Automatic Speech Recognition (ASR) in Burmese .
 
-- **Base Model:** facebook/wav2vec2-xls-r-300m
-- **Language:** Burmese (Myanmar / my)
-- **Architecture:** Acoustic Transformer with Linear CTC Output Head
-- **Target Vocabulary:** 64 character tokens
-- **License:** Apache 2.0
+## Model Overview
 
-## Benchmark Results
+- **Architecture:** Wav2Vec2 XLS-R (300M)
+- **Parameters:** ~317M
+- **Task:** Automatic Speech Recognition (ASR)
+- **Language:** Burmese (`my`)
+- **Sampling Rate:** 16,000 Hz (Mono)
+- **Tokenizer / Vocab:** 64 Burmese Character Tokens (CTC Head)
 
-### Evaluation Progression (Validation Set)
+## Dataset Details
 
-| Step | Training Loss | Validation Loss | CER (%) | WER (%) | chrF |
-| :---: | :---: | :---: | :---: | :---: | :---: |
-| Baseline | - | - | 113.39 | 100.00 | 0.55 |
-| 250 | 13.05 | 1.16 | 17.40 | 72.71 | 66.26 |
-| 500 | 3.28 | 0.42 | 7.33 | 39.69 | 87.39 |
-| 750 | 1.68 | 0.31 | 5.68 | 33.10 | 91.12 |
-| 1000 | 1.12 | 0.29 | 4.68 | 29.82 | 92.75 |
-| 1250 | 0.75 | 0.25 | 4.08 | 27.96 | 93.58 |
-| 1500 | 0.35 | 0.27 | 3.81 | 25.96 | 94.43 |
-| 1750 | 0.13 | 0.31 | 3.67 | 25.18 | 94.70 |
-| 2000 | 0.05 | 0.31 | 3.53 | 24.25 | 94.97 |
+The model was trained on a standardized, high-quality Burmese speech corpus:
 
-### Generalization on Held-Out Unseen Speakers (Test Set)
+- **Total Duration:** ~22 Hours of audio
+- **Total Utterances:** 24,560 WAV files (16 kHz, 16-bit PCM, Mono)
+- **Total Speakers:** 13 Synthetic Burmese Speakers 
+- **Speakers:** ပီယ ၊ ဝါစာ ၊ သီရိ ၊ ဒီပ ၊ အက္ခရာ ၊ သဒ္ဒါ ၊ သရ ၊ တာရာ ၊ ကဝိ ၊ သုတ ၊ ပသာဒ ၊ ဂီတ ၊ နန္ဒ
+- **Training & Validation:** 11 speakers (20,699 train utterances, 2,300 validation utterances)
+- **Test Set (Held-Out):** 2 unseen speakers ( ဂီတ & နန္ဒ , 1,561 utterances )
+- **Text Preprocessing:** Standardized Myanmar Unicode (`\u1000-\u109F`) with unified word segmentation and punctuation removal. The held-out test split evaluates zero-shot acoustic generalization across completely unseen voices.
 
-Evaluated on 1,561 test samples (1.69 hours) from 2 speakers (Gita and Nanda) held out completely from training and validation:
+## Training Configuration
 
-| Metric | Baseline (Untrained) | Fine-Tuned (Step 2000) | Absolute Improvement |
-| :--- | :---: | :---: | :---: |
-| CER (%) | 113.39 | 6.58 | +106.81% |
-| WER (%) | 100.00 | 33.24 | +66.76% |
-| chrF | 0.55 | 87.69 | +87.14 |
-| SER (%) | 100.00 | 91.35 | +8.65% |
-| DER (%) | 85.06 | 5.87 | +79.19% |
-| IER (%) | 0.00 | 2.57 | -2.57% |
+| Parameter | Value |
+| :--- | :--- |
+| Effective Batch Size | 32 (4 per device x 8 gradient accumulation) |
+| Peak Learning Rate | 3.0e-4 |
+| Learning Rate Scheduler | Linear |
+| Warmup Steps | 200 |
+| Total Steps | 2,000 (~6 Epochs) |
+| Mixed Precision | BF16 |
+| Gradient Checkpointing | Enabled (Input Require Grads) |
+| Augmentation | None |
 
-### Comparative Benchmark Across Models
+## Evaluation Results
 
-| Model | Parameters | Training Steps | Val CER (%) | Val WER (%) | Val chrF | Test CER (%) | Test WER (%) | Test chrF |
-| :--- | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: |
-| Whisper Tiny | 39M | 1,500 | 4.47 | 25.21 | 92.91 | 11.69 | 42.29 | 79.82 |
-| Whisper Base | 74M | 2,000 | 4.43 | 26.55 | 91.94 | 8.55 | 36.29 | 83.71 |
-| Wav2Vec2 XLS-R | 317M | 2,000 | 3.53 | 24.25 | 94.97 | 6.58 | 33.24 | 87.69 |
+| Stage / Evaluation | Step | Train Loss | Val Loss | WER (%) | CER (%) | SER (%) | DER (%) | IER (%) | chrF |
+| :--- | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: |
+| Baseline (Untrained) | 0 | - | - | 100.00 | 113.39 | 100.00 | 85.06 | 0.00 | 0.55 |
+| Validation Set | 250 | 13.0478 | 1.1601 | 72.71 | 17.40 | 99.78 | 13.98 | 3.03 | 66.26 |
+| Validation Set | 500 | 3.2769 | 0.4160 | 39.69 | 7.33 | 87.70 | 6.37 | 3.16 | 87.39 |
+| Validation Set | 750 | 1.6812 | 0.3104 | 33.10 | 5.68 | 83.43 | 3.66 | 4.38 | 91.12 |
+| Validation Set | 1000 | 1.1205 | 0.2926 | 29.82 | 4.68 | 78.35 | 5.50 | 2.43 | 92.75 |
+| Validation Set | 1250 | 0.7550 | 0.2558 | 27.96 | 4.08 | 75.09 | 5.15 | 2.70 | 93.58 |
+| Validation Set | 1500 | 0.3550 | 0.2688 | 25.96 | 3.81 | 71.17 | 4.16 | 3.21 | 94.43 |
+| Validation Set | 1750 | 0.1300 | 0.3087 | 25.18 | 3.67 | 70.57 | 4.77 | 2.47 | 94.70 |
+| Validation Set | 2000 | 0.0465 | 0.3125 | 24.25 | 3.53 | 69.13 | 3.77 | 3.07 | 94.97 |
+| **UNSEEN TEST (Final)** | **Final** | **-** | **-** | **33.24** | **6.58** | **91.35** | **5.87** | **2.57** | **87.69** |
 
-## Dataset Specifications
+## Usage
 
-- **Total Audio Files:** 24,560 WAV recordings (22.38 hours)
-- **Audio Format:** 16,000 Hz, 16-bit Mono WAV
-- **Speakers:** 13 native Burmese speakers (5 female, 8 male)
-- **Split Distribution:**
-  - Train: 20,699 samples (18.61 hours)
-  - Validation: 2,300 samples (2.08 hours)
-  - Test: 1,561 samples (1.69 hours) — 2 unseen held-out speakers (Gita & Nanda)
+```python
+import torch
+from transformers import pipeline
 
-## Training Hyperparameters
+pipe = pipeline(
+    'automatic-speech-recognition',
+    model='thantzinphyo/Wav2Vec2-XLS-R-300M-ASR',
+    device='cuda:0' if torch.cuda.is_available() else 'cpu',
+)
 
-- **Base Architecture:** facebook/wav2vec2-xls-r-300m
-- **Max Steps:** 2,000
-- **Effective Batch Size:** 32 (per_device=4, gradient_accumulation_steps=8)
-- **Learning Rate:** 3e-4 (Linear warmup 200 steps)
-- **Precision:** BF16
-- **Gradient Checkpointing:** Enabled (enable_input_require_grads)
+output = pipe('audio.wav')
+print(output['text'])