Datasets:
Add comprehensive README with benchmark findings
Browse files
README.md
CHANGED
|
@@ -1,8 +1,43 @@
|
|
| 1 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 2 |
|
| 3 |
## Overview
|
| 4 |
|
| 5 |
-
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**.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 6 |
|
| 7 |
## Text Normalization
|
| 8 |
|
|
@@ -18,20 +53,6 @@ All WER and CER metrics are computed **after applying full text normalization**
|
|
| 18 |
- And other labialized character normalizations
|
| 19 |
4. **Whitespace Normalization**: Multiple spaces collapsed to single space
|
| 20 |
|
| 21 |
-
## Dataset Description
|
| 22 |
-
|
| 23 |
-
This benchmark evaluates ASR models on various challenging conditions:
|
| 24 |
-
|
| 25 |
-
| Split | Description |
|
| 26 |
-
|-------|-------------|
|
| 27 |
-
| `augmented` | Standard test set with data augmentation |
|
| 28 |
-
| `low_audio` | Low-quality audio samples |
|
| 29 |
-
| `over_augmented` | Heavily augmented audio |
|
| 30 |
-
| `over_augmented_telecom` | Heavily augmented telecom-specific audio |
|
| 31 |
-
| `telecom` | Real telecom customer service recordings |
|
| 32 |
-
| `telecom_random` | Random subset of telecom recordings |
|
| 33 |
-
| `train` | Training set evaluation (for reference) |
|
| 34 |
-
|
| 35 |
## Evaluated Models
|
| 36 |
|
| 37 |
The benchmark includes the following model families:
|
|
@@ -54,17 +75,17 @@ The benchmark includes the following model families:
|
|
| 54 |
- `facebook/mms-1b-all` - Meta's Massively Multilingual Speech
|
| 55 |
- `openai/whisper-small` - OpenAI Whisper Small
|
| 56 |
|
| 57 |
-
## Dataset
|
| 58 |
|
| 59 |
-
This dataset contains three
|
| 60 |
|
| 61 |
-
### 1. `WER`
|
| 62 |
-
Word Error Rate (WER) scores for all models across all
|
| 63 |
|
| 64 |
-
### 2. `CER`
|
| 65 |
-
Character Error Rate (CER) scores for all models across all
|
| 66 |
|
| 67 |
-
### 3. `Gender_GAP`
|
| 68 |
Gender fairness analysis showing:
|
| 69 |
- Male WER/CER scores
|
| 70 |
- Female WER/CER scores
|
|
@@ -72,31 +93,9 @@ Gender fairness analysis showing:
|
|
| 72 |
|
| 73 |
## Key Findings
|
| 74 |
|
| 75 |
-
### Best Performing Models by Split
|
| 76 |
-
|
| 77 |
-
| Split | Best Model | WER (%) |
|
| 78 |
-
|-------|------------|---------|
|
| 79 |
-
| augmented | badrex/Ethio-ASR-amharic | 38.72 |
|
| 80 |
-
| low_audio | omniASR_LLM_3B_v2 | 27.26 |
|
| 81 |
-
| over_augmented | badrex/Ethio-ASR-amharic | 42.08 |
|
| 82 |
-
| over_augmented_telecom | badrex/Ethio-ASR-amharic | 74.74 |
|
| 83 |
-
| telecom | badrex/Ethio-ASR-amharic | 45.13 |
|
| 84 |
-
| telecom_random | badrex/Ethio-ASR-amharic | 46.66 |
|
| 85 |
-
| train | omniASR_LLM_3B_v2 | 27.29 |
|
| 86 |
-
|
| 87 |
### Gender Fairness Analysis
|
| 88 |
|
| 89 |
-
Average WER gender gap (Female - Male) across models
|
| 90 |
-
|
| 91 |
-
| Split | Avg Gap (%) | Interpretation |
|
| 92 |
-
|-------|-------------|----------------|
|
| 93 |
-
| augmented | +3.80 | Female disadvantaged |
|
| 94 |
-
| low_audio | +4.06 | Female disadvantaged |
|
| 95 |
-
| over_augmented | +3.77 | Female disadvantaged |
|
| 96 |
-
| over_augmented_telecom | +13.64 | Female disadvantaged |
|
| 97 |
-
| telecom | +6.89 | Female disadvantaged |
|
| 98 |
-
| telecom_random | +6.62 | Female disadvantaged |
|
| 99 |
-
| train | +4.69 | Female disadvantaged |
|
| 100 |
|
| 101 |
### Key Observations
|
| 102 |
|
|
@@ -114,34 +113,12 @@ Average WER gender gap (Female - Male) across models:
|
|
| 114 |
from datasets import load_dataset
|
| 115 |
|
| 116 |
# Load WER scores
|
| 117 |
-
wer_data = load_dataset("SAARAI/
|
| 118 |
|
| 119 |
# Load CER scores
|
| 120 |
-
cer_data = load_dataset("SAARAI/
|
| 121 |
|
| 122 |
# Load Gender Gap analysis
|
| 123 |
-
gender_data = load_dataset("SAARAI/
|
| 124 |
-
```
|
| 125 |
-
|
| 126 |
-
## Citation
|
| 127 |
-
|
| 128 |
-
If you use this benchmark, please cite:
|
| 129 |
-
|
| 130 |
-
```bibtex
|
| 131 |
-
@misc{ethiotelebench2024,
|
| 132 |
-
title={EthioTelecomBench: A Benchmark for Amharic ASR in Telecom Domain},
|
| 133 |
-
author={SAARAI},
|
| 134 |
-
year={2024},
|
| 135 |
-
publisher={Hugging Face},
|
| 136 |
-
url={https://huggingface.co/datasets/SAARAI/EthioTelecomBench}
|
| 137 |
-
}
|
| 138 |
```
|
| 139 |
|
| 140 |
-
## License
|
| 141 |
-
|
| 142 |
-
This benchmark is released under the Apache 2.0 License.
|
| 143 |
-
|
| 144 |
-
## Links
|
| 145 |
-
|
| 146 |
-
- [Interactive Error Visualization](https://ethio-asr-error-viz.vercel.app/)
|
| 147 |
-
- [Source Repository](https://github.com/IsraelAbebe/Ethio-ASR)
|
|
|
|
| 1 |
+
---
|
| 2 |
+
language:
|
| 3 |
+
- am
|
| 4 |
+
license: apache-2.0
|
| 5 |
+
task_categories:
|
| 6 |
+
- automatic-speech-recognition
|
| 7 |
+
tags:
|
| 8 |
+
- asr
|
| 9 |
+
- amharic
|
| 10 |
+
- ethiopian
|
| 11 |
+
- telecom
|
| 12 |
+
- benchmark
|
| 13 |
+
- wer
|
| 14 |
+
- cer
|
| 15 |
+
- gender-fairness
|
| 16 |
+
- gsma
|
| 17 |
+
pretty_name: GSMA EthioTelecomBench
|
| 18 |
+
size_categories:
|
| 19 |
+
- n<1K
|
| 20 |
+
---
|
| 21 |
+
|
| 22 |
+
# GSMA EthioTelecomBench: Amharic ASR Benchmark for Telecom Domain
|
| 23 |
|
| 24 |
## Overview
|
| 25 |
|
| 26 |
+
GSMA 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**.
|
| 27 |
+
|
| 28 |
+
## Audio Conditions
|
| 29 |
+
|
| 30 |
+
The benchmark evaluates models across different audio conditions:
|
| 31 |
+
|
| 32 |
+
| Column Name | Description |
|
| 33 |
+
|-------------|-------------|
|
| 34 |
+
| Clean Studio | High-quality studio recordings |
|
| 35 |
+
| Ambient Room Noise | Standard recordings with background noise |
|
| 36 |
+
| Harsh Environment | Heavily degraded audio conditions |
|
| 37 |
+
| Distant / Low-Volume Mic | Low-quality microphone recordings |
|
| 38 |
+
| Wideband Phone (VoLTE) | Real telecom VoLTE recordings |
|
| 39 |
+
| Randomized Phone Channel | Random phone channel conditions |
|
| 40 |
+
| Narrowband 2G + Packet Loss | Degraded 2G network with packet loss |
|
| 41 |
|
| 42 |
## Text Normalization
|
| 43 |
|
|
|
|
| 53 |
- And other labialized character normalizations
|
| 54 |
4. **Whitespace Normalization**: Multiple spaces collapsed to single space
|
| 55 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 56 |
## Evaluated Models
|
| 57 |
|
| 58 |
The benchmark includes the following model families:
|
|
|
|
| 75 |
- `facebook/mms-1b-all` - Meta's Massively Multilingual Speech
|
| 76 |
- `openai/whisper-small` - OpenAI Whisper Small
|
| 77 |
|
| 78 |
+
## Dataset Files
|
| 79 |
|
| 80 |
+
This dataset contains three CSV tables:
|
| 81 |
|
| 82 |
+
### 1. `WER.csv`
|
| 83 |
+
Word Error Rate (WER) scores for all models across all audio conditions. Lower is better.
|
| 84 |
|
| 85 |
+
### 2. `CER.csv`
|
| 86 |
+
Character Error Rate (CER) scores for all models across all audio conditions. Lower is better.
|
| 87 |
|
| 88 |
+
### 3. `Gender_GAP.csv`
|
| 89 |
Gender fairness analysis showing:
|
| 90 |
- Male WER/CER scores
|
| 91 |
- Female WER/CER scores
|
|
|
|
| 93 |
|
| 94 |
## Key Findings
|
| 95 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 96 |
### Gender Fairness Analysis
|
| 97 |
|
| 98 |
+
Average WER gender gap (Female - Male) across models shows **female speakers are consistently disadvantaged** across all audio conditions, with gaps ranging from +3.77% to +13.64%.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 99 |
|
| 100 |
### Key Observations
|
| 101 |
|
|
|
|
| 113 |
from datasets import load_dataset
|
| 114 |
|
| 115 |
# Load WER scores
|
| 116 |
+
wer_data = load_dataset("SAARAI/GSMAEthioTelecomBench", data_files="data/WER.csv", split="train")
|
| 117 |
|
| 118 |
# Load CER scores
|
| 119 |
+
cer_data = load_dataset("SAARAI/GSMAEthioTelecomBench", data_files="data/CER.csv", split="train")
|
| 120 |
|
| 121 |
# Load Gender Gap analysis
|
| 122 |
+
gender_data = load_dataset("SAARAI/GSMAEthioTelecomBench", data_files="data/Gender_GAP.csv", split="train")
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 123 |
```
|
| 124 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|