SaarAI commited on
Commit
28c1311
·
verified ·
1 Parent(s): 81c0c36

Add comprehensive README with benchmark findings

Browse files
Files changed (1) hide show
  1. README.md +48 -71
README.md CHANGED
@@ -1,8 +1,43 @@
1
- # EthioTelecomBench: Amharic ASR Benchmark for Telecom Domain
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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 Splits
58
 
59
- This dataset contains three tables as different splits:
60
 
61
- ### 1. `WER` Split
62
- Word Error Rate (WER) scores for all models across all evaluation splits. Lower is better.
63
 
64
- ### 2. `CER` Split
65
- Character Error Rate (CER) scores for all models across all evaluation splits. Lower is better.
66
 
67
- ### 3. `Gender_GAP` Split
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/EthioTelecomBench", split="WER")
118
 
119
  # Load CER scores
120
- cer_data = load_dataset("SAARAI/EthioTelecomBench", split="CER")
121
 
122
  # Load Gender Gap analysis
123
- gender_data = load_dataset("SAARAI/EthioTelecomBench", split="Gender_GAP")
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