ATOM: Armenian Tiny Optimized Model

A compact, morpheme-aware Automatic Speech Recognition (ASR) model that significantly outperforms OpenAI's Whisper on Armenian speech recognition.

Model Description

ATOM is a specialized ASR model for low-resource Armenian, achieving 64.5% lower WER than vanilla Whisper-tiny on Armenian while using 28% fewer parameters. The model combines:

• Frozen Whisper-tiny encoder (pre-trained audio feature extraction)
• Custom compact decoder (2 layers, trained from scratch on Armenian)
• Morpheme-level BPE tokenization (5,000 tokens optimized for Armenian morphology vs Whisper's 51k multilingual tokens)

Architecture

Input: Audio (16kHz) 
  ↓
Whisper Encoder (frozen, 4 layers, 384 hidden, 1536 FFN)
  ↓
Compact Decoder (trainable, 2 layers, 384 hidden, 1024 FFN)
  ↓
Morpheme Vocabulary (5,000 tokens)
  ↓
Output: Armenian Text

Total Parameters: ~28M (28% smaller than Whisper-tiny's 39M)

Performance

Evaluated on Common Voice 20.0 Armenian test set:

*Whisper WER values for Armenian from published benchmarks

Key Insights:

• ATOM outperforms ALL Whisper models on Armenian, including models up to 55× larger
• Word Error Rate (WER): 42.1% vs Whisper-tiny's 118.6% on Armenian
• Model Size: 28M parameters (28% smaller than Whisper-tiny, 55× smaller than Whisper-large-v2)
• Training Efficiency: Trained on minimal Armenian speech data vs Whisper's 680k hours multilingual

Note: While Whisper models achieve strong performance on high-resource languages (e.g., Whisper-tiny: 79.0% average WER), they perform significantly worse on low-resource Armenian (118.6% WER), demonstrating the need for language-specific approaches.

Why ATOM Outperforms Whisper

1. Morpheme-Aware Tokenization: Armenian is an agglutinative language where words combine multiple morphemes (e.g., "չէինք" = "չ" [negation] + "է" [to be] + "ինք" [we/past]). ATOM's morpheme-level vocabulary (5k tokens) captures this linguistic structure better than Whisper's multilingual word-level BPE (51k tokens).

2. Language-Specific Training: While Whisper is trained on 99 languages (680k hours), ATOM's decoder is trained exclusively on Armenian, allowing deep specialization on Armenian phonology and morphology.

3. Efficient Architecture: The compact 2-layer decoder prevents overfitting on limited training data while the frozen pre-trained encoder provides robust audio feature extraction.

4. Low-Resource Optimization: Whisper's multilingual training spreads capacity across languages, disadvantaging low-resource Armenian. ATOM dedicates all decoder capacity to Armenian.

Intended Uses

Primary Uses:

• Armenian speech-to-text transcription
• Real-time subtitling for Armenian content
• Accessibility tools for Armenian speakers
• Research on morpheme-aware ASR for agglutinative languages

Best Performance:

• Clear speech in quiet environments
• Native Armenian speakers
• Standard Eastern/Western Armenian dialects

Limitations

• Trained on limited data (relatively small dataset)
• May struggle with heavy accents or noisy audio
• Optimized for Armenian only (not multilingual)
• 10% exact match rate indicates room for improvement in perfect transcriptions
• Performance may degrade on out-of-domain audio (non-Common Voice data)

Training Details

Training Data

• Dataset: Common Voice 20.0 Armenian
• Splits Used: Train + Other
• Duration: Approximately 30 hours of Armenian speech
• Speakers: 400+ unique speakers
• Demographics:
  • Gender: 55% Female, 25% Male, 20% Undefined
  • Age: Primarily 20s-30s (70%+)
• Test Set: Common Voice test split (separate, unseen data)

Training Hyperparameters

learning_rate: 1e-4
train_batch_size: 32
gradient_accumulation_steps: 1
warmup_steps: 500
max_steps: 12,000
save_steps: 3,000
fp16: True
optimizer: AdamW (torch)
lr_scheduler_type: cosine
max_grad_norm: 1.0
gradient_checkpointing: True
dataloader_num_workers: 8

Training Infrastructure

• GPU: NVIDIA RTX 3060 ti with FP16 mixed precision
• Framework:
  • Transformers 4.56.2
  • PyTorch 2.8.0+cu129
  • Datasets 3.5.0
  • Tokenizers 0.22.1
• Final Checkpoint: Step 9,000
• Evaluation Loss: 1.36

Usage

Installation

pip install transformers torch torchaudio

Basic Inference

from transformers import WhisperForConditionalGeneration, WhisperProcessor
import torch

# Load model and processor
model = WhisperForConditionalGeneration.from_pretrained("Chillarmo/ATOM")
processor = WhisperProcessor.from_pretrained("Chillarmo/ATOM")

# Load audio (16kHz)
import torchaudio
audio, sr = torchaudio.load("audio.wav")
if sr != 16000:
    resampler = torchaudio.transforms.Resample(sr, 16000)
    audio = resampler(audio)

# Process
input_features = processor(
    audio.squeeze().numpy(), 
    sampling_rate=16000, 
    return_tensors="pt"
).input_features

# Generate
with torch.no_grad():
    predicted_ids = model.generate(
        input_features,
        max_length=448,
        num_beams=5,
        repetition_penalty=1.2,
        no_repeat_ngram_size=3
    )

# Decode
transcription = processor.batch_decode(predicted_ids, skip_special_tokens=True)[0]
print(transcription)

Advanced Usage with Pipeline

from transformers import pipeline

# Create ASR pipeline
asr_pipeline = pipeline(
    "automatic-speech-recognition",
    model="Chillarmo/ATOM",
    device=0  # Use GPU if available
)

# Transcribe
result = asr_pipeline(
    "audio.wav",
    generate_kwargs={
        "max_length": 448,
        "num_beams": 5,
        "repetition_penalty": 1.2
    }
)
print(result["text"])

Technical Details

Morpheme Tokenization

The model uses a custom BPE tokenizer trained on Armenian text with morpheme-level granularity:

• Vocabulary Size: 5,000 tokens
• Special Tokens: <pad>, <s>, </s>, <unk>
• Training Corpus: Armenian Wikipedia + Common Voice transcriptions
• Morpheme Segmentation: Whitespace pre-tokenization optimized for Armenian word structure

Example tokenization:

Word: "չէինք" (we were not)
Morphemes: ["չ", "է", "ինք"]
Translation: [negation] + [to be] + [we/past]

Model Architecture

Encoder (Frozen):

• 4 Transformer encoder layers
• 384 hidden dimensions
• 1536 feed-forward dimensions
• 6 attention heads
• Pre-trained on Whisper's 680k hour multilingual dataset

Decoder (Trained from Scratch):

• 2 Transformer decoder layers (50% reduction)
• 384 hidden dimensions
• 1024 feed-forward dimensions (33% reduction)
• 6 attention heads
• Trained exclusively on Armenian

Reproduction

To reproduce training:

# Install dependencies
pip install transformers datasets evaluate jiwer accelerate

# Train
python train.py \
  --model_name_or_path openai/whisper-tiny \
  --dataset Chillarmo/common_voice_20_armenian \
  --output_dir ./atom-model \
  --learning_rate 1e-4 \
  --per_device_train_batch_size 32 \
  --max_steps 12000 \
  --fp16 \
  --save_steps 3000

Citation

@misc{movsesyan2025atom,
  title={ATOM: Morpheme-Aware Whisper for Low-Resource Armenian ASR},
  author={Movsesyan, Movses},
  year={2025},
  institution={California State University, Sacramento}
}

References

Whisper Armenian benchmarks from published evaluations on Common Voice datasets.

Acknowledgments

• Built on OpenAI's Whisper architecture (Radford et al., 2022)
• Trained on Mozilla Common Voice data
• Morpheme tokenization inspired by Armenian linguistic structure
• California State University, Sacramento

License

[Specify license - typically MIT or Apache 2.0]

---

Model Card Contact: movsesmovsesyan@csus.edu