# EthioTelecomBench: Amharic ASR Benchmark for Telecom Domain

## Overview

EthioTelecomBench is a comprehensive benchmark for evaluating Automatic Speech Recognition (ASR) systems on Amharic speech, with a focus on telecom customer service conversations. This dataset contains evaluation results from **12 models** across **7 evaluation splits**.

## Text Normalization

All WER and CER metrics are computed **after applying full text normalization** to both reference and predicted transcriptions:

1. **Number-to-Text Conversion**: Arabic numerals are converted to Amharic text (e.g., "123" → "አንድ መቶ ሃያ ሶስት")
2. **Punctuation Removal**: All punctuation marks (including Ge'ez punctuation ፠፡።፣፤፥፦፧፨) are removed
3. **Character Normalization**: Ge'ez character variants are normalized to canonical forms:
   - ሀ/ሐ/ኅ/ኻ/ኃ → ሃ
   - ሠ/ሡ/ሢ/ሣ/ሤ/ሥ/ሦ → ሰ/ሱ/ሲ/ሳ/ሴ/ስ/ሶ
   - ዐ/ዑ/ዒ/ዓ/ዔ/ዕ/ዖ → አ/ኡ/ኢ/አ/ኤ/እ/ኦ
   - ፀ/ፁ/ፂ/ፃ/ፄ/ፅ/ፆ → ጸ/ጹ/ጺ/ጻ/ጼ/ጽ/ጾ
   - And other labialized character normalizations
4. **Whitespace Normalization**: Multiple spaces collapsed to single space

## Dataset Description

This benchmark evaluates ASR models on various challenging conditions:

| Split | Description |
|-------|-------------|
| `augmented` | Standard test set with data augmentation |
| `low_audio` | Low-quality audio samples |
| `over_augmented` | Heavily augmented audio |
| `over_augmented_telecom` | Heavily augmented telecom-specific audio |
| `telecom` | Real telecom customer service recordings |
| `telecom_random` | Random subset of telecom recordings |
| `train` | Training set evaluation (for reference) |

## Evaluated Models

The benchmark includes the following model families:

### Ethio-ASR Models (badrex)
- `badrex/Ethio-ASR-amharic` - Amharic-specific model
- `badrex/Ethio-ASR-multilingual-94M` - 94M parameter multilingual
- `badrex/Ethio-ASR-multilingual-300M` - 300M parameter multilingual
- `badrex/Ethio-ASR-multilingual-600M` - 600M parameter multilingual

### OmniASR Models
- `omniASR_CTC_300M_v2` - CTC-based 300M
- `omniASR_CTC_1B_v2` - CTC-based 1B
- `omniASR_CTC_3B_v2` - CTC-based 3B
- `omniASR_LLM_300M_v2` - LLM-based 300M
- `omniASR_LLM_1B_v2` - LLM-based 1B
- `omniASR_LLM_3B_v2` - LLM-based 3B

### Baseline Models
- `facebook/mms-1b-all` - Meta's Massively Multilingual Speech
- `openai/whisper-small` - OpenAI Whisper Small

## Dataset Splits

This dataset contains three tables as different splits:

### 1. `WER` Split
Word Error Rate (WER) scores for all models across all evaluation splits. Lower is better.

### 2. `CER` Split  
Character Error Rate (CER) scores for all models across all evaluation splits. Lower is better.

### 3. `Gender_GAP` Split
Gender fairness analysis showing:
- Male WER/CER scores
- Female WER/CER scores
- Gender gap (Female - Male): Positive values indicate higher error rates for female speakers

## Key Findings

### Best Performing Models by Split

| Split | Best Model | WER (%) |
|-------|------------|---------|
| augmented | badrex/Ethio-ASR-amharic | 38.72 |
| low_audio | omniASR_LLM_3B_v2 | 27.26 |
| over_augmented | badrex/Ethio-ASR-amharic | 42.08 |
| over_augmented_telecom | badrex/Ethio-ASR-amharic | 74.74 |
| telecom | badrex/Ethio-ASR-amharic | 45.13 |
| telecom_random | badrex/Ethio-ASR-amharic | 46.66 |
| train | omniASR_LLM_3B_v2 | 27.29 |

### Gender Fairness Analysis

Average WER gender gap (Female - Male) across models:

| Split | Avg Gap (%) | Interpretation |
|-------|-------------|----------------|
| augmented | +3.80 | Female disadvantaged |
| low_audio | +4.06 | Female disadvantaged |
| over_augmented | +3.77 | Female disadvantaged |
| over_augmented_telecom | +13.64 | Female disadvantaged |
| telecom | +6.89 | Female disadvantaged |
| telecom_random | +6.62 | Female disadvantaged |
| train | +4.69 | Female disadvantaged |

### Key Observations

1. **Domain Adaptation Matters**: Models fine-tuned on Amharic (Ethio-ASR family) significantly outperform general multilingual models on telecom domain data.

2. **Gender Bias**: Most models show higher error rates for female speakers, indicating a systematic gender bias in ASR performance.

3. **Audio Quality Impact**: Performance degrades significantly on low-quality and over-augmented audio, highlighting the need for robust ASR systems.

4. **Model Size vs Performance**: Larger models (3B parameters) generally perform better, but domain-specific smaller models can be competitive.

## Usage

```python
from datasets import load_dataset

# Load WER scores
wer_data = load_dataset("SAARAI/EthioTelecomBench", split="WER")

# Load CER scores
cer_data = load_dataset("SAARAI/EthioTelecomBench", split="CER")

# Load Gender Gap analysis
gender_data = load_dataset("SAARAI/EthioTelecomBench", split="Gender_GAP")
```

## Citation

If you use this benchmark, please cite:

```bibtex
@misc{ethiotelebench2024,
  title={EthioTelecomBench: A Benchmark for Amharic ASR in Telecom Domain},
  author={SAARAI},
  year={2024},
  publisher={Hugging Face},
  url={https://huggingface.co/datasets/SAARAI/EthioTelecomBench}
}
```

## License

This benchmark is released under the Apache 2.0 License.

## Links

- [Interactive Error Visualization](https://ethio-asr-error-viz.vercel.app/)
- [Source Repository](https://github.com/IsraelAbebe/Ethio-ASR)