Whisper Large v3 Turbo Armenian: High-Performance Armenian Speech Recognition
This model is a fine-tuned version of openai/whisper-large-v3-turbo on the Chillarmo/common_voice_20_armenian dataset. This model represents a significant advancement in Armenian automatic speech recognition, achieving state-of-the-art performance with substantially improved accuracy compared to smaller variants.
Model Details
Model Description
This is a fine-tuned Whisper Large v3 Turbo model specifically optimized for Armenian speech recognition. The model leverages the latest Whisper architecture improvements while maintaining the efficiency of the Turbo variant, providing an optimal balance between performance and computational requirements for Armenian language processing.
- Developed by: Movses Movsesyan (Independent Research)
- Model type: Automatic Speech Recognition
- Language(s): Armenian (hy)
- License: Apache 2.0
- Finetuned from model: openai/whisper-large-v3-turbo
Model Sources
- Repository: Hugging Face Model Hub
- Base Model: OpenAI Whisper
- Paper: Robust Speech Recognition via Large-Scale Weak Supervision
Performance Highlights
🚀 Exceptional Accuracy: Achieves a 15.31% WER and 2.86% CER - representing significant improvements over smaller models:
| Model | WER | CER | Exact Match |
|---|---|---|---|
| Whisper Large v3 Turbo Armenian | 15.31% | 2.86% | 42.73% |
| Whisper Small Armenian v2 | 24.01% | 4.77% | 28.14% |
Key Improvements:
- 36% reduction in Word Error Rate compared to the small model
- 40% reduction in Character Error Rate
- 52% improvement in Exact Match accuracy
- Superior performance while maintaining efficient inference speed
Uses
Direct Use
This model excels at transcribing Armenian speech to text with high accuracy, making it suitable for:
- Production-grade Armenian speech transcription systems
- Real-time Armenian voice interfaces with minimal errors
- Professional Armenian media content processing
- High-accuracy Armenian voice assistants and applications
- Academic and research applications in Armenian computational linguistics
Downstream Use
The model can be integrated into enterprise and research applications such as:
- Professional Armenian voice assistants and chatbots
- High-quality subtitle generation for Armenian media
- Accessibility tools requiring high transcription accuracy
- Educational platforms for Armenian language learning
- Call center analytics and voice processing systems
Out-of-Scope Use
This model should not be used for:
- Speech recognition in languages other than Armenian
- Speaker identification or verification tasks
- Audio classification beyond speech transcription
- Critical applications requiring 100% accuracy (medical/legal without human review)
Bias, Risks, and Limitations
While this model achieves excellent performance, users should be aware of:
- Domain specificity: Performance may vary across different speaking styles and domains
- Audio quality dependency: Optimal performance requires reasonably clear audio input
- Dialectal variations: Performance may vary across different Armenian dialects
- Computational requirements: Larger model size requires more computational resources than smaller variants
Recommendations
For optimal results:
- Test thoroughly on your specific use case and audio conditions
- Implement appropriate error handling and confidence thresholds
- Consider computational requirements when deploying at scale
- Monitor performance across different speaker demographics and accents
How to Get Started with the Model
Use the code below to get started with the model:
from transformers import AutoProcessor, AutoModelForSpeechSeq2Seq
import torch
# Load the processor and model
processor = AutoProcessor.from_pretrained("Chillarmo/whisper-large-v3-turbo-armenian")
model = AutoModelForSpeechSeq2Seq.from_pretrained("Chillarmo/whisper-large-v3-turbo-armenian")
# Enable half precision for faster inference (optional)
model = model.half()
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model.to(device)
def transcribe_armenian(audio_path):
"""
Transcribe Armenian audio file to text
Args:
audio_path (str): Path to audio file
Returns:
str: Transcribed text
"""
import librosa
# Load and process audio file
audio, sr = librosa.load(audio_path, sr=16000)
# Process the audio
inputs = processor(audio, sampling_rate=16000, return_tensors="pt")
inputs = inputs.to(device)
# Generate transcription
with torch.no_grad():
predicted_ids = model.generate(
inputs["input_features"],
max_new_tokens=448,
do_sample=False,
use_cache=True
)
# Decode the transcription
transcription = processor.batch_decode(predicted_ids, skip_special_tokens=True)
return transcription[0]
# Example usage
# transcription = transcribe_armenian("path/to/armenian_audio.wav")
# print(f"Transcription: {transcription}")
Batch Processing Example
def transcribe_batch(audio_files):
"""
Transcribe multiple audio files efficiently
Args:
audio_files (list): List of audio file paths
Returns:
list: List of transcriptions
"""
transcriptions = []
for audio_file in audio_files:
transcription = transcribe_armenian(audio_file)
transcriptions.append(transcription)
return transcriptions
# Example batch processing
# audio_files = ["audio1.wav", "audio2.wav", "audio3.wav"]
# results = transcribe_batch(audio_files)
Training Details
Training Data
The model was fine-tuned on the Chillarmo/common_voice_20_armenian dataset, utilizing the same high-quality Armenian speech data that has proven effective for Armenian speech recognition tasks.
Training Procedure
Training Hyperparameters
The following hyperparameters were used during training:
- Training regime: Mixed precision training with optimized settings for Large v3 Turbo
- Epochs: 2.65 (optimal convergence achieved)
- Training runtime: 48,521 seconds (approximately 13.5 hours)
- Training samples per second: 1.649
- Training steps per second: 0.103
- Final training loss: 0.106
- Total training steps: 5,000
Performance Metrics During Training
- Final evaluation loss: 0.069
- Training efficiency: Optimized for the Large v3 Turbo architecture
- Convergence: Achieved excellent performance within 2.65 epochs
Evaluation
Testing Data, Factors & Metrics
Testing Data
The model was evaluated on a held-out test set from the Chillarmo/common_voice_20_armenian dataset using the same evaluation methodology as the baseline models for fair comparison.
Metrics
The model was evaluated using standard speech recognition metrics:
- Word Error Rate (WER): Percentage of words incorrectly transcribed
- Character Error Rate (CER): Percentage of characters incorrectly transcribed
- Exact Match: Percentage of utterances transcribed perfectly
Results
The fine-tuned model achieved exceptional performance on the evaluation set:
| Metric | Value | Improvement vs Small Model |
|---|---|---|
| Word Error Rate (WER) | 15.31% | -36.2% (24.01% → 15.31%) |
| Character Error Rate (CER) | 2.86% | -40.0% (4.77% → 2.86%) |
| Exact Match | 42.73% | +51.8% (28.14% → 42.73%) |
| Average Prediction Length | 7.76 tokens | Consistent with ground truth |
| Average Label Length | 7.77 tokens | - |
| Length Ratio | 0.999 | Excellent length calibration |
Performance Analysis
The results demonstrate exceptional performance characteristics:
- Superior accuracy: Significant improvements across all metrics compared to smaller models
- Length calibration: Near-perfect length ratio (0.999) indicates excellent model calibration
- Consistency: High exact match rate (42.73%) shows the model frequently produces perfect transcriptions
- Robustness: Low character error rate (2.86%) indicates strong character-level understanding
Technical Specifications
Model Architecture and Objective
This model is based on the Whisper Large v3 Turbo architecture, featuring:
- Encoder: Advanced Transformer encoder processing mel-spectrogram features
- Decoder: Optimized Transformer decoder for efficient text generation
- Architecture: Enhanced Transformer sequence-to-sequence model
- Model size: Large v3 Turbo (optimized for efficiency)
- Input: 128-dimensional log mel-spectrograms
- Output: High-accuracy Armenian text transcriptions
- Special features: Turbo optimizations for faster inference while maintaining quality
Compute Infrastructure
Hardware
Training was performed on the following hardware configuration:
- Training duration: ~13.5 hours for optimal convergence
- Computational efficiency: Optimized training pipeline for Large v3 Turbo architecture
- Memory optimization: Efficient memory usage during training and inference
Software
- Framework: Hugging Face Transformers (latest version)
- Training library: PyTorch with advanced optimization
- Audio processing: librosa, soundfile
- Evaluation: datasets, evaluate, jiwer
- Optimization: Mixed precision training for efficiency
Deployment Considerations
Hardware Requirements
Minimum Requirements:
- GPU: 8GB VRAM (for optimal performance)
- CPU: Modern multi-core processor
- RAM: 16GB system memory
Recommended for Production:
- GPU: 16GB+ VRAM for batch processing
- CPU: High-performance multi-core processor
- RAM: 32GB+ for large-scale deployment
Performance Optimization Tips
- Use half precision (
model.half()) for faster inference - Batch processing for multiple audio files
- GPU acceleration strongly recommended
- Caching enabled for repeated inference tasks
Citation
BibTeX:
@misc{movsesyan2025whisper-large-v3-turbo-armenian,
author = {Movsesyan, Movses},
title = {Whisper Large v3 Turbo Armenian},
year = {2025},
publisher = {Hugging Face},
url = {https://huggingface.co/Chillarmo/whisper-large-v3-turbo-armenian}
}
@article{radford2022robust,
title={Robust speech recognition via large-scale weak supervision},
author={Radford, Alec and Kim, Jong Wook and Xu, Tao and Brockman, Greg and McLeavey, Christine and Sutskever, Ilya},
journal={International Conference on Machine Learning},
pages={28492--28518},
year={2023},
organization={PMLR}
}
APA:
Movsesyan, M. (2025). Whisper Large v3 Turbo Armenian. Hugging Face. https://huggingface.co/Chillarmo/whisper-large-v3-turbo-armenian
Model Card Authors
This model card was created by Movses Movsesyan based on the fine-tuning results and comprehensive performance evaluation of the Whisper Large v3 Turbo Armenian model.
Acknowledgments
Special thanks to the OpenAI team for the Whisper architecture and the Common Voice project contributors for providing high-quality Armenian speech data that made this research possible.