thantzinphyo/whisper-small-myanmar-phase1

🇲🇲 Whisper Small Myanmar (Burmese) ASR - Phase 1

This model is a fine-tuned version of openai/whisper-small (244 Million parameters) dedicated for Myanmar (Burmese) speech recognition.

Benchmark Results (Phase 1)

Metric	Baseline (Zero-Shot)	Fine-Tuned (Phase 1)	Net Improvement
CER (Character Error Rate)	159.97%	17.94%	-142.03%
WER (Word Error Rate)	317.49%	56.31%	-261.18%
Sentence Exact Match	0.00%	2.06%	+2.06%
Evaluation Loss	3.3531	1.5270	-1.8261

Model & Training Architecture

• Base Architecture: openai/whisper-small (244M parameters, 12 encoder layers, 12 decoder layers)
• Strategy: Partial Encoder Freezing (Layers 0-7 frozen, Layers 8-11 + all 12 decoder layers trainable)
• Tonal Language Safeguards: Phase-vocoder speed perturbation (0.9x, 1.0x, 1.1x) + F0-protected SpecAugment (protecting 80-300Hz fundamental tone frequencies)
• Optimization: AdamW + Cosine Annealing Learning Rate Scheduler (Peak LR: 3.75e-5, 1000 Steps)
• Language Force Token: Burmese (<|my|>) with task='transcribe' and Unicode NFC normalization

Usage (Inference Pipeline)

import torch
from transformers import pipeline

# Load pipeline directly from HuggingFace
pipe = pipeline(
    "automatic-speech-recognition",
    model="thantzinphyo/whisper-small-myanmar-phase1",
    torch_dtype=torch.float16,
    device="cuda:0" if torch.cuda.is_available() else "cpu"
)

# Transcribe Burmese audio (16kHz WAV)
result = pipe("your_burmese_audio.wav", generate_kwargs={"language": "my", "task": "transcribe"})
print("Transcription:", result["text"])