docs: Update model card
Browse filesThis changes a handful of stylistic matters:
1. Collapse some sections to make the model card more digestible
2. Introduce the conversational performance results first, as these are the primary distinguishing scores. Showing the read-aloud scores first would give the reader the impression that there's no difference between v1 and v2
In the conversational performance results, we currently only compare the model with other Røst models. We should really give the conversational performance the same treatment as the read-aloud one. This would include comparisons with Whisper-v3-large and Hviske-v2 in the associated table and plots.
README.md
CHANGED
|
@@ -30,10 +30,13 @@ model-index:
|
|
| 30 |
name: WER
|
| 31 |
---
|
| 32 |
|
| 33 |
-
#
|
| 34 |
-
This is a
|
|
|
|
|
|
|
|
|
|
|
|
|
| 35 |
|
| 36 |
-
This repository contains a Wav2Vec2 model trained on the [CoRal-v2 dataset](https://huggingface.co/datasets/CoRal-project/coral-v2/tree/main). The CoRal-v2 dataset includes a rich variety of Danish conversational and read-aloud data, distributed across diverse age groups, genders, and dialects. The model is designed for automatic speech recognition (ASR).
|
| 37 |
|
| 38 |
## Quick Start
|
| 39 |
|
|
@@ -59,233 +62,270 @@ Next you can use the model using the `transformers` Python package as follows:
|
|
| 59 |
|
| 60 |
Explore the following audio samples along with their transcriptions and accuracy metrics. Each example showcases the model's performance with different Danish dialects.
|
| 61 |
|
| 62 |
-
|
| 63 |
-
|
| 64 |
-
|
| 65 |
-
|
| 66 |
-
|
| 67 |
-
|
| 68 |
-
|
| 69 |
-
|
| 70 |
-
|
| 71 |
-
|
| 72 |
-
|
| 73 |
-
**
|
| 74 |
-
*det blev til
|
| 75 |
-
|
| 76 |
-
|
| 77 |
-
|
| 78 |
-
|
| 79 |
-
|
| 80 |
-
|
| 81 |
-
|
| 82 |
-
|
| 83 |
-
|
| 84 |
-
<
|
| 85 |
-
|
| 86 |
-
|
| 87 |
-
|
| 88 |
-
|
| 89 |
-
|
| 90 |
-
|
| 91 |
-
|
| 92 |
-
|
| 93 |
-
|
| 94 |
-
|
| 95 |
-
|
| 96 |
-
|
| 97 |
-
|
| 98 |
-
|
| 99 |
-
|
| 100 |
-
|
| 101 |
-
|
| 102 |
-
|
| 103 |
-
|
| 104 |
-
|
| 105 |
-
|
| 106 |
-
</
|
| 107 |
-
|
| 108 |
-
|
| 109 |
-
|
| 110 |
-
|
| 111 |
-
|
| 112 |
-
|
| 113 |
-
|
| 114 |
-
|
| 115 |
-
|
| 116 |
-
|
| 117 |
-
|
| 118 |
-
|
| 119 |
-
|
| 120 |
-
|
| 121 |
-
**
|
| 122 |
-
|
| 123 |
-
|
| 124 |
-
|
| 125 |
-
|
| 126 |
-
|
| 127 |
-
|
| 128 |
-
|
| 129 |
-
|
| 130 |
-
**
|
| 131 |
-
|
| 132 |
-
|
| 133 |
-
|
| 134 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 135 |
|
| 136 |
---
|
| 137 |
|
| 138 |
## Model Details
|
| 139 |
|
| 140 |
Wav2Vec2 is a state-of-the-art model architecture for speech recognition, leveraging self-supervised learning from raw audio data. The pre-trained [Wav2Vec2-XLS-R-300M](https://huggingface.co/facebook/wav2vec2-xls-r-300m) has been fine-tuned for automatic speech recognition with the [CoRal-v2 dataset](https://huggingface.co/datasets/CoRal-project/coral-v2/tree/main) dataset to enhance its performance in recognizing Danish speech with consideration to different dialects. The model was trained for 30K steps using the training setup in the [CoRaL repository](https://github.com/alexandrainst/coral/tree) by running:
|
| 141 |
-
```
|
| 142 |
-
python src/scripts/finetune_asr_model.py model=wav2vec2-small max_steps=30000 datasets.coral_conversation_internal.id=CoRal-project/coral-v2 datasets.coral_readaloud_internal.id=CoRal-project/coral-v2
|
| 143 |
-
```
|
| 144 |
-
The model is evaluated using a Language Model (LM) as post-processing. The utilized LM is the one trained and used by [CoRal-project/roest-wav2vec2-315m-v1](https://huggingface.co/CoRal-project/roest-wav2vec2-315m-v2).
|
| 145 |
|
| 146 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 147 |
|
| 148 |
-
|
|
|
|
| 149 |
|
| 150 |
-
|
| 151 |
-
|
| 152 |
-
|
| 153 |
-
- Read-aloud
|
| 154 |
-
- **Language**: Danish.
|
| 155 |
-
- **Variation**: Includes various dialects, age groups, and gender distinctions.
|
| 156 |
-
### License
|
| 157 |
-
Note that the dataset used is licensed under a custom license, adapted from OpenRAIL-M, which allows commercial use with a few restrictions (speech synthesis and biometric identification). See [license](https://huggingface.co/Alvenir/coral-1-whisper-large/blob/main/LICENSE).
|
| 158 |
|
| 159 |
---
|
| 160 |
|
| 161 |
## Evaluation
|
| 162 |
|
| 163 |
The model was evaluated using the following metrics:
|
| 164 |
-
- **Word Error Rate (WER)**: The percentage of words incorrectly transcribed.
|
| 165 |
- **Character Error Rate (CER)**: The percentage of characters incorrectly transcribed.
|
|
|
|
| 166 |
|
| 167 |
-
|
| 168 |
-
|
| 169 |
-
|
| 170 |
-
| Model | Number of parameters | Finetuned on data of type | [CoRal](https://huggingface.co/datasets/CoRal-project/coral/viewer/read_aloud/test) CER | [CoRal](https://huggingface.co/datasets/CoRal-project/coral/viewer/read_aloud/test) WER |
|
| 171 |
-
| :----------------------------------------------------------------------------------------------- | -------------------: | --------------------------: | --------------------------------------------------------------------------------------: | --------------------------------------------------------------------------------------: |
|
| 172 |
-
| [CoRal-project/roest-wav2vec2-1B-v2](https://huggingface.co/CoRal-project/roest-wav2vec2-1B-v2) | 1B | Read-aloud and conversation | 6.5% ± 0.2% | 16.4% ± 0.4% |
|
| 173 |
-
| [CoRal-project/roest-wav2vec2-315M-v2](https://huggingface.co/CoRal-project/roest-wav2vec2-315m-v2) | 315M | Read-aloud and conversation | 6.5% ± 0.2% | 16.3% ± 0.4% |
|
| 174 |
-
| [CoRal-project/roest-whisper-large-v1](https://huggingface.co/CoRal-project/roest-whisper-large-v1) | 1540M | Read-aloud | **4.3% ± 0.2%** | **10.4% ± 0.3%** |
|
| 175 |
-
| [CoRal-project/roest-wav2vec2-315M-v1](https://huggingface.co/CoRal-project/roest-wav2vec2-315M-v1) | 315M | Read-aloud | 6.6% ± 0.2% | 17.0% ± 0.4% |
|
| 176 |
-
| [mhenrichsen/hviske-v2](https://huggingface.co/syvai/hviske-v2) | 1540M | Read-aloud | 4.7% ± 0.2% | 11.8% ± 0.3% |
|
| 177 |
-
| [openai/whisper-large-v3](https://hf.co/openai/whisper-large-v3) | 1540M | - | 11.4% ± 0.3% | 28.3% ± 0.6% |
|
| 178 |
|
| 179 |
-
|
| 180 |
|
| 181 |
-
The
|
|
|
|
|
|
|
|
|
|
| 182 |
|
| 183 |
| Model | Number of parameters | Finetuned on data of type | [CoRal-v2::conversation](https://huggingface.co/datasets/CoRal-project/coral-v2/viewer/conversation/test) CER | [CoRal-v2::conversation](https://huggingface.co/datasets/CoRal-project/coral-v2/viewer/conversation/test) WER |
|
| 184 |
| :-------------------------------------------------------------------------------------------------- | -------------------: | --------------------------: | ------------------------------------------------------------------------------------------------------------: | ------------------------------------------------------------------------------------------------------------: |
|
| 185 |
-
| [CoRal-project/roest-wav2vec2-1B-v2](https://huggingface.co/CoRal-project/roest-wav2vec2-1B-v2) | 1B | Read-aloud and conversation |
|
| 186 |
| [CoRal-project/roest-wav2vec2-315M-v2](https://huggingface.co/CoRal-project/roest-wav2vec2-315m-v2) | 315M | Read-aloud and conversation | 24.2% | 37.7% |
|
| 187 |
-
| [CoRal-project/roest-whisper-large-v1](https://huggingface.co/CoRal-project/roest-whisper-large-v1)
|
| 188 |
-
| [CoRal-project/roest-wav2vec2-315m-v1](https://huggingface.co/CoRal-project/roest-wav2vec2-315m-v1)
|
| 189 |
|
| 190 |
|
| 191 |
-
###
|
| 192 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 193 |
|
| 194 |
<img src="https://huggingface.co/CoRal-project/roest-wav2vec2-315m-v2/resolve/main/images/cer.png">
|
| 195 |
|
| 196 |
-
|
| 197 |
-
|
| 198 |
-
|
| 199 |
-
|
| 200 |
-
|
| 201 |
-
|
| 202 |
-
|
| 203 |
-
|
| 204 |
-
|
|
| 205 |
-
|
| 206 |
-
|
|
| 207 |
-
|
|
| 208 |
-
|
|
| 209 |
-
|
|
| 210 |
-
|
|
| 211 |
-
|
|
| 212 |
-
|
|
| 213 |
-
|
|
| 214 |
-
|
|
| 215 |
-
|
| 216 |
-
|
| 217 |
-
|
| 218 |
-
|
|
| 219 |
-
|
| 220 |
-
|
|
| 221 |
-
|
|
| 222 |
-
|
| 223 |
-
|
| 224 |
-
|
| 225 |
-
|
| 226 |
-
|
| 227 |
-
|
| 228 |
-
|
| 229 |
-
|
| 230 |
-
|
|
| 231 |
-
|
| 232 |
-
|
|
| 233 |
-
|
|
| 234 |
-
|
|
| 235 |
-
|
|
| 236 |
-
|
| 237 |
-
|
| 238 |
-
|
| 239 |
-
|
| 240 |
-
|
| 241 |
-
|
| 242 |
-
|
|
| 243 |
-
|
|
| 244 |
-
|
|
| 245 |
-
|
|
| 246 |
-
|
|
| 247 |
-
|
|
| 248 |
-
|
| 249 |
-
|
| 250 |
-
|
| 251 |
-
|
| 252 |
-
|
| 253 |
-
|
| 254 |
-
|
| 255 |
-
|
| 256 |
-
|
| 257 |
-
|
| 258 |
-
|
| 259 |
-
|
| 260 |
-
|
| 261 |
-
|
| 262 |
-
|
|
| 263 |
-
|
|
| 264 |
-
|
|
| 265 |
-
|
|
| 266 |
-
|
|
| 267 |
-
|
| 268 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 269 |
|
| 270 |
### Performance on Other Datasets
|
| 271 |
|
| 272 |
The model was also tested against other datasets to evaluate generalizability:
|
| 273 |
|
| 274 |
-
| | **
|
| 275 |
-
| ------------------------------------------------------------------------------------- | -------------------------- |
|
| 276 |
-
| Evaluation Dataset
|
| 277 |
-
| [CoRal](https://huggingface.co/datasets/CoRal-project/coral/viewer/read_aloud/test) | **10.4** | **4.3**
|
| 278 |
-
| [NST-da](https://huggingface.co/datasets/alexandrainst/nst-da) | 29.8 | 14.5
|
| 279 |
-
| [CommonVoice17](https://huggingface.co/datasets/mozilla-foundation/common_voice_17_0) | 15.6 | 8.2
|
| 280 |
-
| [Fleurs-da_dk](https://huggingface.co/datasets/google/fleurs) | **12.6** | **5.1**
|
| 281 |
|
| 282 |
**OBS!** The vocab used for training incudes numerals (0,1,2,..,9), which are translated to text in a post-processing step. If the model misses spaces the numbers are interpreted as one, which especially affects the NST score as this dataset contains many numerals.
|
| 283 |
|
| 284 |
---
|
| 285 |
-
### Note on comparing whisper and wav2vec2 models
|
| 286 |
-
The Whisper models detailed in this model card exhibit significantly lower Character Error Rates (CER) and Word Error Rates (WER) compared to the Wav2Vec2 models. Whisper utilizes a transformer-based architecture with additional layers that enhance contextual understanding. In contrast, Wav2Vec2 models employ shorter context windows that focus on sound prediction. The Roest-Wav2Vec2 models incorporate a straightforward language model during post-processing, which addresses errors based on statistical language patterns. Introducing a more complex, contextual post-processing language model might enable a better comparison between these model types, which the CoRal project plans to explore in future releases.
|
| 287 |
|
| 288 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 289 |
|
| 290 |
---
|
| 291 |
|
|
@@ -307,12 +347,13 @@ The CoRal project is funded by the [Danish Innovation Fund](https://innovationsf
|
|
| 307 |
|
| 308 |
We would like specifically to thank Dan Saattrup Nielsen, Alexandra Institute for (among other things) the repository work and Simon Leminen Madsen, Alexandra Institute for modelling work.
|
| 309 |
|
|
|
|
| 310 |
## Citation
|
| 311 |
|
| 312 |
```bibtex
|
| 313 |
@misc{roest-wav2vec2-315m-v2,
|
| 314 |
author = {Marie Juhl Jørgensen, Søren Vejlgaard Holm, Martin Carsten Nielsen, Dan Saattrup Nielsen, Sif Bernstorff Lehmann, Simon Leminen Madsen and Torben Blach},
|
| 315 |
-
title = {
|
| 316 |
year = {2025},
|
| 317 |
url = {https://huggingface.co/CoRal-project/roest-wav2vec2-315m-v2},
|
| 318 |
}
|
|
|
|
| 30 |
name: WER
|
| 31 |
---
|
| 32 |
|
| 33 |
+
# Røst-wav2vec2-315m-v2
|
| 34 |
+
This is a Danish state-of-the-art speech recognition model, trained as part of the CoRal project by [Alvenir](https://www.alvenir.ai/).
|
| 35 |
+
|
| 36 |
+
This repository contains a Wav2Vec2 model trained on the [CoRal-v2 dataset](https://huggingface.co/datasets/CoRal-project/coral-v2/tree/main).
|
| 37 |
+
The CoRal-v2 dataset includes a rich variety of Danish conversational and read-aloud data, distributed across diverse age groups, genders, and dialects.
|
| 38 |
+
The model is designed for automatic speech recognition (ASR).
|
| 39 |
|
|
|
|
| 40 |
|
| 41 |
## Quick Start
|
| 42 |
|
|
|
|
| 62 |
|
| 63 |
Explore the following audio samples along with their transcriptions and accuracy metrics. Each example showcases the model's performance with different Danish dialects.
|
| 64 |
|
| 65 |
+
<details>
|
| 66 |
+
<summary>
|
| 67 |
+
<b>Example 1 - Vestjysk Dialect</b>
|
| 68 |
+
</summary>
|
| 69 |
+
|
| 70 |
+
**Audio Sample:**
|
| 71 |
+
<audio controls>
|
| 72 |
+
<source src="https://huggingface.co/CoRal-project/roest-wav2vec2-315m-v2/resolve/main/audio_samples/example1.wav" type="audio/wav">
|
| 73 |
+
Your browser does not support the audio tag.
|
| 74 |
+
</audio>
|
| 75 |
+
|
| 76 |
+
**Model Transcription:**
|
| 77 |
+
*det blev til yderlig ti mål i den første sæson på trods af en position som back*
|
| 78 |
+
|
| 79 |
+
**Target Transcription:**
|
| 80 |
+
*det blev til yderligere ti mål i den første sæson på trods af en position som back*
|
| 81 |
+
|
| 82 |
+
- **Character Error Rate (CER):** 3.7%
|
| 83 |
+
- **Word Error Rate (WER):** 5.9%
|
| 84 |
+
</details>
|
| 85 |
+
|
| 86 |
+
<details>
|
| 87 |
+
<summary>
|
| 88 |
+
<b>Example 2 - Sønderjysk Dialect</b>
|
| 89 |
+
</summary>
|
| 90 |
+
|
| 91 |
+
**Audio Sample:**
|
| 92 |
+
<audio controls>
|
| 93 |
+
<source src="https://huggingface.co/CoRal-project/roest-wav2vec2-315m-v2/resolve/main/audio_samples/example2.wav" type="audio/wav">
|
| 94 |
+
Your browser does not support the audio tag.
|
| 95 |
+
</audio>
|
| 96 |
+
|
| 97 |
+
**Model Transcription:**
|
| 98 |
+
*en arkitektoniske udformning af pladser forslagene iver benzen*
|
| 99 |
+
|
| 100 |
+
**Target Transcription:**
|
| 101 |
+
*den arkitektoniske udformning af pladsen er forestået af ivar bentsen*
|
| 102 |
+
|
| 103 |
+
- **Character Error Rate (CER):** 20.3%
|
| 104 |
+
- **Word Error Rate (WER):** 60.0%
|
| 105 |
+
</details>
|
| 106 |
+
|
| 107 |
+
<details>
|
| 108 |
+
<summary>
|
| 109 |
+
<b>Example 3 - Nordsjællandsk Dialect</b>
|
| 110 |
+
</summary>
|
| 111 |
+
|
| 112 |
+
**Audio Sample:**
|
| 113 |
+
<audio controls>
|
| 114 |
+
<source src="https://huggingface.co/CoRal-project/roest-wav2vec2-315m-v2/resolve/main/audio_samples/example3.wav" type="audio/wav">
|
| 115 |
+
Your browser does not support the audio tag.
|
| 116 |
+
</audio>
|
| 117 |
+
|
| 118 |
+
**Model Transcription:**
|
| 119 |
+
*østrig og ungarn samarbejder om søen gennem den østrigske og ungarske vandkommission*
|
| 120 |
+
|
| 121 |
+
**Target Transcription:**
|
| 122 |
+
*østrig og ungarn samarbejder om søen gennem den østrigske og ungarske vandkommission*
|
| 123 |
+
|
| 124 |
+
- **Character Error Rate (CER):** 0.0%
|
| 125 |
+
- **Word Error Rate (WER):** 0.0%
|
| 126 |
+
</details>
|
| 127 |
+
|
| 128 |
+
<details>
|
| 129 |
+
<summary>
|
| 130 |
+
<b>Example 4 - Lollandsk Dialect</b>
|
| 131 |
+
</summary>
|
| 132 |
+
|
| 133 |
+
**Audio Sample:**
|
| 134 |
+
<audio controls>
|
| 135 |
+
<source src="https://huggingface.co/CoRal-project/roest-wav2vec2-315m-v2/resolve/main/audio_samples/example4.wav" type="audio/wav">
|
| 136 |
+
Your browser does not support the audio tag.
|
| 137 |
+
</audio>
|
| 138 |
+
|
| 139 |
+
**Model Transcription:**
|
| 140 |
+
*det er produceret af thomas helme og indspillede i easy sound recording studio i københavn*
|
| 141 |
+
|
| 142 |
+
**Target Transcription:**
|
| 143 |
+
*det er produceret af thomas helmig og indspillet i easy sound recording studio i københavn*
|
| 144 |
+
|
| 145 |
+
- **Character Error Rate (CER):** 4.4%
|
| 146 |
+
- **Word Error Rate (WER):** 13.3%
|
| 147 |
+
</details>
|
| 148 |
|
| 149 |
---
|
| 150 |
|
| 151 |
## Model Details
|
| 152 |
|
| 153 |
Wav2Vec2 is a state-of-the-art model architecture for speech recognition, leveraging self-supervised learning from raw audio data. The pre-trained [Wav2Vec2-XLS-R-300M](https://huggingface.co/facebook/wav2vec2-xls-r-300m) has been fine-tuned for automatic speech recognition with the [CoRal-v2 dataset](https://huggingface.co/datasets/CoRal-project/coral-v2/tree/main) dataset to enhance its performance in recognizing Danish speech with consideration to different dialects. The model was trained for 30K steps using the training setup in the [CoRaL repository](https://github.com/alexandrainst/coral/tree) by running:
|
|
|
|
|
|
|
|
|
|
|
|
|
| 154 |
|
| 155 |
+
```bash
|
| 156 |
+
python src/scripts/finetune_asr_model.py \
|
| 157 |
+
model=wav2vec2-small \
|
| 158 |
+
max_steps=30000 \
|
| 159 |
+
datasets.coral_conversation_internal.id=CoRal-project/coral-v2 \
|
| 160 |
+
datasets.coral_readaloud_internal.id=CoRal-project/coral-v2
|
| 161 |
+
```
|
| 162 |
|
| 163 |
+
The model is evaluated using a Language Model (LM) as post-processing.
|
| 164 |
+
The utilized LM is the one trained and used by [CoRal-project/roest-wav2vec2-315m-v1](https://huggingface.co/CoRal-project/roest-wav2vec2-315m-v1).
|
| 165 |
|
| 166 |
+
The model was trained on the [CoRal-v2](https://huggingface.co/datasets/CoRal-project/coral-v2/tree/main) dataset, including both the conversational and read-aloud subset.
|
| 167 |
+
This dataset consists of Danish speech across a variety of dialects, age groups and gender distinctions.
|
| 168 |
+
Note that the dataset, and thus also this model, is licensed under a custom license, adapted from OpenRAIL-M, which allows commercial use with few restrictions (speech synthesis and biometric identification) - see [license](https://huggingface.co/Alvenir/coral-1-whisper-large/blob/main/LICENSE).
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 169 |
|
| 170 |
---
|
| 171 |
|
| 172 |
## Evaluation
|
| 173 |
|
| 174 |
The model was evaluated using the following metrics:
|
|
|
|
| 175 |
- **Character Error Rate (CER)**: The percentage of characters incorrectly transcribed.
|
| 176 |
+
- **Word Error Rate (WER)**: The percentage of words incorrectly transcribed.
|
| 177 |
|
| 178 |
+
### Conversational CoRal Performance
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 179 |
|
| 180 |
+
The model was firstly evaluated on a tentative version of the coral-v2 conversation dataset.
|
| 181 |
|
| 182 |
+
The results are tentative as the test set only includes 5 unique speakers, of which 4 are women.
|
| 183 |
+
The test set includes 2 speakers with 'Fynsk' dialect, 1 with 'Sønderjysk', 1 with 'Non-native' and 1 'Nordjysk'.
|
| 184 |
+
The Whisper model is performing very poorly on the test set. An explanation could be hallucinations during silence and short sentences, a known whisper issue.
|
| 185 |
+
Furthermore, both v1 models have not been trained on any conversation data, giving the models an obvious disadvantage.
|
| 186 |
|
| 187 |
| Model | Number of parameters | Finetuned on data of type | [CoRal-v2::conversation](https://huggingface.co/datasets/CoRal-project/coral-v2/viewer/conversation/test) CER | [CoRal-v2::conversation](https://huggingface.co/datasets/CoRal-project/coral-v2/viewer/conversation/test) WER |
|
| 188 |
| :-------------------------------------------------------------------------------------------------- | -------------------: | --------------------------: | ------------------------------------------------------------------------------------------------------------: | ------------------------------------------------------------------------------------------------------------: |
|
| 189 |
+
| [CoRal-project/roest-wav2vec2-1B-v2](https://huggingface.co/CoRal-project/roest-wav2vec2-1B-v2) | 1B | Read-aloud and conversation | **23.9%** | **36.7%** |
|
| 190 |
| [CoRal-project/roest-wav2vec2-315M-v2](https://huggingface.co/CoRal-project/roest-wav2vec2-315m-v2) | 315M | Read-aloud and conversation | 24.2% | 37.7% |
|
| 191 |
+
| [CoRal-project/roest-whisper-large-v1](https://huggingface.co/CoRal-project/roest-whisper-large-v1) | 1540M | Read-aloud | 138% | 121% |
|
| 192 |
+
| [CoRal-project/roest-wav2vec2-315m-v1](https://huggingface.co/CoRal-project/roest-wav2vec2-315m-v1) | 315M | Read-aloud | 123% | 80.5% |
|
| 193 |
|
| 194 |
|
| 195 |
+
### Read-aloud CoRal Performance
|
| 196 |
+
|
| 197 |
+
| Model | Number of parameters | Finetuned on data of type | [CoRal](https://huggingface.co/datasets/CoRal-project/coral/viewer/read_aloud/test) CER | [CoRal](https://huggingface.co/datasets/CoRal-project/coral/viewer/read_aloud/test) WER |
|
| 198 |
+
| :-------------------------------------------------------------------------------------------------- | -------------------: | --------------------------: | --------------------------------------------------------------------------------------: | --------------------------------------------------------------------------------------: |
|
| 199 |
+
| [CoRal-project/roest-wav2vec2-1B-v2](https://huggingface.co/CoRal-project/roest-wav2vec2-1B-v2) | 1B | Read-aloud and conversation | 6.5% ± 0.2% | 16.4% ± 0.4% |
|
| 200 |
+
| [CoRal-project/roest-wav2vec2-315M-v2](https://huggingface.co/CoRal-project/roest-wav2vec2-315m-v2) | 315M | Read-aloud and conversation | 6.5% ± 0.2% | 16.3% ± 0.4% |
|
| 201 |
+
| [CoRal-project/roest-whisper-large-v1](https://huggingface.co/CoRal-project/roest-whisper-large-v1) | 1540M | Read-aloud | **4.3% ± 0.2%** | **10.4% ± 0.3%** |
|
| 202 |
+
| [CoRal-project/roest-wav2vec2-315M-v1](https://huggingface.co/CoRal-project/roest-wav2vec2-315M-v1) | 315M | Read-aloud | 6.6% ± 0.2% | 17.0% ± 0.4% |
|
| 203 |
+
| [mhenrichsen/hviske-v2](https://huggingface.co/syvai/hviske-v2) | 1540M | Read-aloud | 4.7% ± 0.2% | 11.8% ± 0.3% |
|
| 204 |
+
| [openai/whisper-large-v3](https://hf.co/openai/whisper-large-v3) | 1540M | - | 11.4% ± 0.3% | 28.3% ± 0.6% |
|
| 205 |
+
|
| 206 |
+
**OBS!** Benchmark for hviske-v2 has been re-evaluated and the confidence interval is larger than reported in the model card.
|
| 207 |
|
| 208 |
<img src="https://huggingface.co/CoRal-project/roest-wav2vec2-315m-v2/resolve/main/images/cer.png">
|
| 209 |
|
| 210 |
+
<img src="https://huggingface.co/CoRal-project/roest-wav2vec2-315m-v2/resolve/main/images/wer.png">
|
| 211 |
+
|
| 212 |
+
|
| 213 |
+
<details>
|
| 214 |
+
<summary>
|
| 215 |
+
<b>Detailed CER scores in % of evaluation across demographics on the CoRal test data</b>
|
| 216 |
+
</summary>
|
| 217 |
+
|
| 218 |
+
| Category | Røst-whisper-large-v1 | Røst-wav2vec2-315m-v1 | Røst-wav2vec2-315m-v2 | Røst-wav2vec2-1B-v2 |
|
| 219 |
+
|:---:|:---:|:---:|:---:|:---:|
|
| 220 |
+
| female | 5.1 | 7.4 | 7.2 | 7.3 |
|
| 221 |
+
| male | 3.6 | 5.8 | 5.7 | 5.8 |
|
| 222 |
+
| 0-25 | 3.4 | 5.4 | 5.3 | 5.1 |
|
| 223 |
+
| 25-50 | 4.0 | 6.2 | 6.0 | 5.7 |
|
| 224 |
+
| 50+ | 5.0 | 7.5 | 7.4 | 7.8 |
|
| 225 |
+
| Bornholmsk | 3.8 | 6.8 | 6.1 | 6.2 |
|
| 226 |
+
| Fynsk | 5.1 | 7.4 | 7.2 | 6.9 |
|
| 227 |
+
| Københavnsk | 1.9 | 3.3 | 3.2 | 3.0 |
|
| 228 |
+
| Non-native | 4.8 | 7.8 | 7.5 | 7.3 |
|
| 229 |
+
| Nordjysk | 1.6 | 2.6 | 2.8 | 2.6 |
|
| 230 |
+
| Sjællandsk | 3.0 | 4.4 | 4.5 | 3.9 |
|
| 231 |
+
| Sydømål | 4.1 | 6.4 | 6.4 | 6.5 |
|
| 232 |
+
| Sønderjysk | 8.8 | 11.9 | 11.6 | 12.6 |
|
| 233 |
+
| Vestjysk | 6.4 | 10.1 | 9.8 | 10.5 |
|
| 234 |
+
| Østjysk | 2.6 | 4.0 | 4.1 | 3.8 |
|
| 235 |
+
| Overall | 4.3 | 6.6 | 6.5 | 6.5 |
|
| 236 |
+
|
| 237 |
+
</details>
|
| 238 |
+
|
| 239 |
+
<details>
|
| 240 |
+
<summary>
|
| 241 |
+
<b>Detailed WER scores in % of evaluation across demographics on the CoRal test data</b>
|
| 242 |
+
</summary>
|
| 243 |
+
|
| 244 |
+
| Category | Røst-whisper-large-v1 | Røst-wav2vec2-315m-v1 | Røst-wav2vec2-315m-v2 | Røst-wav2vec2-1B-v2 |
|
| 245 |
+
|:---:|:---:|:---:|:---:|:---:|
|
| 246 |
+
| female | 11.5 | 18.5 | 17.7 | 17.8 |
|
| 247 |
+
| male | 9.4 | 15.5 | 14.9 | 15.0 |
|
| 248 |
+
| 0-25 | 9.0 | 14.7 | 14.0 | 13.7 |
|
| 249 |
+
| 25-50 | 10.1 | 16.6 | 15.8 | 15.3 |
|
| 250 |
+
| 50+ | 11.3 | 18.2 | 17.7 | 18.5 |
|
| 251 |
+
| Bornholmsk | 9.8 | 17.7 | 15.7 | 16.4 |
|
| 252 |
+
| Fynsk | 12.1 | 18.3 | 17.7 | 16.7 |
|
| 253 |
+
| Københavnsk | 5.9 | 10.2 | 10.0 | 9.5 |
|
| 254 |
+
| Non-native | 12.2 | 20.9 | 19.4 | 19.4 |
|
| 255 |
+
| Nordjysk | 4.5 | 7.7 | 7.5 | 7.3 |
|
| 256 |
+
| Sjællandsk | 7.6 | 12.6 | 12.7 | 11.0 |
|
| 257 |
+
| Sydømål | 10.0 | 14.9 | 15.3 | 14.4 |
|
| 258 |
+
| Sønderjysk | 17.5 | 26.0 | 25.4 | 27.8 |
|
| 259 |
+
| Vestjysk | 15.0 | 26.3 | 25.2 | 26.7 |
|
| 260 |
+
| Østjysk | 7.5 | 11.7 | 11.3 | 10.8 |
|
| 261 |
+
| Overall | 10.4 | 17.0 | 16.3 | 16.4 |
|
| 262 |
+
|
| 263 |
+
</details>
|
| 264 |
+
|
| 265 |
+
<details>
|
| 266 |
+
<summary>
|
| 267 |
+
<b>Experiments with Røst-wav2vec2-315M with and without language model</b>
|
| 268 |
+
</summary>
|
| 269 |
+
|
| 270 |
+
The inclusion of a post-processing language model can affect the performance significantly.
|
| 271 |
+
The Røst-v1 and Røst-v2 models are using the same Language Model (LM).
|
| 272 |
+
The utilized LM is the one trained and used by [CoRal-project/roest-wav2vec2-315m-v1](https://huggingface.co/CoRal-project/roest-wav2vec2-315m-v1).
|
| 273 |
+
|
| 274 |
+
| Model | Number of parameters | Finetuned on data of type | Postprocessed with Language Model | [CoRal](https://huggingface.co/datasets/alexandrainst/coral/viewer/read_aloud/test) CER | [CoRal](https://huggingface.com/datasets/alexandrainst/coral/viewer/read_aloud/test) WER |
|
| 275 |
+
| :-------------------------------------------------------------------------------------------------- | -------------------: | --------------------------: | --------------------------------: | --------------------------------------------------------------------------------------: | ---------------------------------------------------------------------------------------: |
|
| 276 |
+
| [CoRal-project/roest-wav2vec2-1B-v2](https://huggingface.co/CoRal-project/roest-wav2vec2-1B-v2) | 1B | Read-aloud and conversation | Yes | **6.5% ± 0.2%** | **16.4% ± 0.4%** |
|
| 277 |
+
| [CoRal-project/roest-wav2vec2-1B-v2](https://huggingface.co/CoRal-project/roest-wav2vec2-1B-v2) | 1B | Read-aloud and conversation | No | 8.1% ± 0.2% | 23.9% ± 0.4% |
|
| 278 |
+
| [CoRal-project/roest-wav2vec2-315M-v2](https://huggingface.co/CoRal-project/roest-wav2vec2-315m-v2) | 315M | Read-aloud and conversation | Yes | **6.5% ± 0.2%** | **16.3% ± 0.4%** |
|
| 279 |
+
| [CoRal-project/roest-wav2vec2-315M-v2](https://huggingface.co/CoRal-project/roest-wav2vec2-315m-v2) | 315M | Read-aloud and conversation | No | 8.2% ± 0.2% | 25.1% ± 0.4% |
|
| 280 |
+
| [CoRal-project/roest-wav2vec2-315m-v1](https://huggingface.co/CoRal-project/roest-wav2vec2-315m-v1) | 315M | Read-aloud | Yes | 6.6% ± 0.2% | 17.0% ± 0.4% |
|
| 281 |
+
| [CoRal-project/roest-wav2vec2-315m-v1](https://huggingface.co/CoRal-project/roest-wav2vec2-315m-v1) | 315M | Read-aloud | No | 8.6% ± 0.2% | 26.3% ± 0.5% |
|
| 282 |
+
|
| 283 |
+
Here are the results of the model on different Danish dialects in the test set:
|
| 284 |
+
|
| 285 |
+
| | Røst-v1 | | Røst-v1 | | Røst-v2 | | Røst-v2 | |
|
| 286 |
+
|-------------|---------|---------|---------|---------|---------|---------|---------|---------|
|
| 287 |
+
| **LM** | **No** | | **Yes** | | **No** | | **Yes** | |
|
| 288 |
+
|-------------|---------|---------|---------|---------|---------|---------|---------|---------|
|
| 289 |
+
| Dialect | CER (%) | WER (%) | CER (%) | WER (%) | CER (%) | WER (%) | CER (%) | WER (%) |
|
| 290 |
+
| Vestjysk | 12.7 | 37.1 | 10.1 | 26.3 | 12.2 | 36.3 | 9.82 | 25.2 |
|
| 291 |
+
| Sønderjysk | 14.7 | 37.8 | 11.9 | 26.0 | 14.2 | 36.2 | 11.6 | 25.4 |
|
| 292 |
+
| Bornholmsk | 9.32 | 29.9 | 6.79 | 17.7 | 8.08 | 26.7 | 6.12 | 15.7 |
|
| 293 |
+
| Østjysk | 5.51 | 18.7 | 3.97 | 11.7 | 5.39 | 18.0 | 4.06 | 11.3 |
|
| 294 |
+
| Nordjysk | 3.86 | 13.6 | 2.57 | 7.72 | 3.80 | 13.5 | 2.75 | 7.51 |
|
| 295 |
+
| Københavnsk | 5.27 | 18.8 | 3.31 | 10.2 | 5.02 | 17.7 | 3.20 | 9.98 |
|
| 296 |
+
| Fynsk | 9.41 | 28.6 | 7.43 | 18.3 | 8.86 | 27.0 | 7.20 | 17.7 |
|
| 297 |
+
| Non-native | 10.6 | 33.2 | 7.84 | 20.9 | 10.0 | 31.6 | 7.46 | 19.4 |
|
| 298 |
+
| Sjællandsk | 5.82 | 19.5 | 4.44 | 12.6 | 5.70 | 18.6 | 4.48 | 12.7 |
|
| 299 |
+
| Sydømål | 7.09 | 20.7 | 6.38 | 14.9 | 6.96 | 20.4 | 6.44 | 15.3 |
|
| 300 |
+
|
| 301 |
+
</details>
|
| 302 |
|
| 303 |
### Performance on Other Datasets
|
| 304 |
|
| 305 |
The model was also tested against other datasets to evaluate generalizability:
|
| 306 |
|
| 307 |
+
| | **Røst-whisper-large-v1** | | **Røst-wav2vec2-315M-v1** | | **Røst-wav2vec2-315M-v2** | | **Røst-wav2vec2-1B-v2** | |
|
| 308 |
+
| ------------------------------------------------------------------------------------- | -------------------------- | --------- | -------------------------- | --------- | -------------------------- | ----------- | ------------------------ | --------- |
|
| 309 |
+
| **Evaluation Dataset** | **WER %** | **CER %** | **WER %** | **CER %** | **WER %** | **CER %** | **WER %** | **CER %** |
|
| 310 |
+
| [CoRal](https://huggingface.co/datasets/CoRal-project/coral/viewer/read_aloud/test) | **10.4** | **4.3** | 17.0 | 6.6 | **16.3** | **6.5** | 16.4 | **6.5** |
|
| 311 |
+
| [NST-da](https://huggingface.co/datasets/alexandrainst/nst-da) | 29.8 | 14.5 | 29.7 | 13.9 | 26.1 | 11.9 | **12.4** | **4.9** |
|
| 312 |
+
| [CommonVoice17](https://huggingface.co/datasets/mozilla-foundation/common_voice_17_0) | 15.6 | 8.2 | 16.7 | 6.6 | **14.4** | **5.4** | 26.3 | 10.9 |
|
| 313 |
+
| [Fleurs-da_dk](https://huggingface.co/datasets/google/fleurs) | **12.6** | **5.1** | 16.6 | 6.3 | 15.6 | 6.1 | **13.7** | **5.5** |
|
| 314 |
|
| 315 |
**OBS!** The vocab used for training incudes numerals (0,1,2,..,9), which are translated to text in a post-processing step. If the model misses spaces the numbers are interpreted as one, which especially affects the NST score as this dataset contains many numerals.
|
| 316 |
|
| 317 |
---
|
|
|
|
|
|
|
| 318 |
|
| 319 |
+
### Note on comparing Whisper and Wav2Vec2 models
|
| 320 |
+
The Whisper models detailed in this model card exhibit significantly lower Character Error Rates (CER) and Word Error Rates (WER) compared to the Wav2Vec2 models.
|
| 321 |
+
Whisper utilizes a transformer-based architecture with additional layers that enhance contextual understanding.
|
| 322 |
+
In contrast, Wav2Vec2 models employ shorter context windows that focus on sound prediction.
|
| 323 |
+
The Røst-Wav2Vec2 models incorporate a straightforward language model during post-processing, which addresses errors based on statistical language patterns.
|
| 324 |
+
Introducing a more complex, contextual post-processing language model might enable a better comparison between these model types, which the CoRal project plans to explore in future releases.
|
| 325 |
+
|
| 326 |
+
The Røst-Whisper model excels in read-aloud data, leveraging its embedded contextual framework to achieve more robust recognition within this context.
|
| 327 |
+
However, Wav2Vec2 models appear to generalize more effectively across various speech recognition tasks, whereas Whisper models incur higher error rates in conversational data.
|
| 328 |
+
It’s important to note that the CoRal-v2 conversation dataset, being tentative and featuring limited speaker diversity, might influence these results.
|
| 329 |
|
| 330 |
---
|
| 331 |
|
|
|
|
| 347 |
|
| 348 |
We would like specifically to thank Dan Saattrup Nielsen, Alexandra Institute for (among other things) the repository work and Simon Leminen Madsen, Alexandra Institute for modelling work.
|
| 349 |
|
| 350 |
+
|
| 351 |
## Citation
|
| 352 |
|
| 353 |
```bibtex
|
| 354 |
@misc{roest-wav2vec2-315m-v2,
|
| 355 |
author = {Marie Juhl Jørgensen, Søren Vejlgaard Holm, Martin Carsten Nielsen, Dan Saattrup Nielsen, Sif Bernstorff Lehmann, Simon Leminen Madsen and Torben Blach},
|
| 356 |
+
title = {Røst-wav2vec-315m-v2: A Danish state-of-the-art speech recognition model trained on varied demographics and dialects},
|
| 357 |
year = {2025},
|
| 358 |
url = {https://huggingface.co/CoRal-project/roest-wav2vec2-315m-v2},
|
| 359 |
}
|