Datasets:
Formats:
parquet
Languages:
English
Size:
1K - 10K
ArXiv:
Tags:
shouts
emotional_speech
distance_speech
smartphone_recordings
nonsense_phrases
non-native_accents
License:
Update README.md
Browse files
README.md
CHANGED
|
@@ -56,7 +56,19 @@ We provide a test, train and validation split
|
|
| 56 |
There is no speaker cross-over between splits, the train and validation sets each contain 10 speakers not seen in the training set
|
| 57 |
|
| 58 |
## Baseline Results
|
| 59 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 60 |
|
| 61 |
## Metadata Details
|
| 62 |
|
|
|
|
| 56 |
There is no speaker cross-over between splits, the train and validation sets each contain 10 speakers not seen in the training set
|
| 57 |
|
| 58 |
## Baseline Results
|
| 59 |
+
|
| 60 |
+
Automatic speech recognition: word error rate, character error rate, phone error rate
|
| 61 |
+
|
| 62 |
+
| Model | WER ↓ | CER ↓ | PER ↓ |
|
| 63 |
+
|----------------------------------|---------|---------|---------|
|
| 64 |
+
| Whisper - medium.en | **17.27%** | 7.81% | 7.80% |
|
| 65 |
+
| Whisper - turbo | 17.93% | **7.28%** | **7.30%** |
|
| 66 |
+
| NeMo Quartznet | 39.49% | 15.24% | 15.77% |
|
| 67 |
+
| NeMo Fastconformer Transducer | 24.96% | 10.72% | 10.13% |
|
| 68 |
+
| Wav2Vec2-Base-960h | 49.65% | 18.94% | 19.90% |
|
| 69 |
+
|
| 70 |
+
|
| 71 |
+
|
| 72 |
|
| 73 |
## Metadata Details
|
| 74 |
|