README.md
| 1 | --- |
| 2 | language: |
| 3 | - multilingual |
| 4 | - ab |
| 5 | - af |
| 6 | - am |
| 7 | - ar |
| 8 | - as |
| 9 | - az |
| 10 | - ba |
| 11 | - be |
| 12 | - bg |
| 13 | - bi |
| 14 | - bo |
| 15 | - br |
| 16 | - bs |
| 17 | - ca |
| 18 | - ceb |
| 19 | - cs |
| 20 | - cy |
| 21 | - da |
| 22 | - de |
| 23 | - el |
| 24 | - en |
| 25 | - eo |
| 26 | - es |
| 27 | - et |
| 28 | - eu |
| 29 | - fa |
| 30 | - fi |
| 31 | - fo |
| 32 | - fr |
| 33 | - gl |
| 34 | - gn |
| 35 | - gu |
| 36 | - gv |
| 37 | - ha |
| 38 | - haw |
| 39 | - hi |
| 40 | - hr |
| 41 | - ht |
| 42 | - hu |
| 43 | - hy |
| 44 | - ia |
| 45 | - id |
| 46 | - is |
| 47 | - it |
| 48 | - he |
| 49 | - ja |
| 50 | - jv |
| 51 | - ka |
| 52 | - kk |
| 53 | - km |
| 54 | - kn |
| 55 | - ko |
| 56 | - la |
| 57 | - lm |
| 58 | - ln |
| 59 | - lo |
| 60 | - lt |
| 61 | - lv |
| 62 | - mg |
| 63 | - mi |
| 64 | - mk |
| 65 | - ml |
| 66 | - mn |
| 67 | - mr |
| 68 | - ms |
| 69 | - mt |
| 70 | - my |
| 71 | - ne |
| 72 | - nl |
| 73 | - nn |
| 74 | - no |
| 75 | - oc |
| 76 | - pa |
| 77 | - pl |
| 78 | - ps |
| 79 | - pt |
| 80 | - ro |
| 81 | - ru |
| 82 | - sa |
| 83 | - sco |
| 84 | - sd |
| 85 | - si |
| 86 | - sk |
| 87 | - sl |
| 88 | - sn |
| 89 | - so |
| 90 | - sq |
| 91 | - sr |
| 92 | - su |
| 93 | - sv |
| 94 | - sw |
| 95 | - ta |
| 96 | - te |
| 97 | - tg |
| 98 | - th |
| 99 | - tk |
| 100 | - tl |
| 101 | - tr |
| 102 | - tt |
| 103 | - uk |
| 104 | - ud |
| 105 | - uz |
| 106 | - vi |
| 107 | - war |
| 108 | - yi |
| 109 | - yo |
| 110 | - zh |
| 111 | thumbnail: |
| 112 | tags: |
| 113 | - audio-classification |
| 114 | - speechbrain |
| 115 | - embeddings |
| 116 | - Language |
| 117 | - Identification |
| 118 | - pytorch |
| 119 | - ECAPA-TDNN |
| 120 | - TDNN |
| 121 | - VoxLingua107 |
| 122 | license: "apache-2.0" |
| 123 | datasets: |
| 124 | - VoxLingua107 |
| 125 | metrics: |
| 126 | - Accuracy |
| 127 | widget: |
| 128 | - example_title: English Sample |
| 129 | src: https://cdn-media.huggingface.co/speech_samples/LibriSpeech_61-70968-0000.flac |
| 130 | --- |
| 131 | |
| 132 | # VoxLingua107 ECAPA-TDNN Spoken Language Identification Model |
| 133 | |
| 134 | ## Model description |
| 135 | |
| 136 | This is a spoken language recognition model trained on the [VoxLingua107 dataset](https://cs.taltech.ee/staff/tanel.alumae/data/voxlingua107/) using SpeechBrain. |
| 137 | The model uses the ECAPA-TDNN architecture that has previously been used for speaker recognition. However, it uses |
| 138 | more fully connected hidden layers after the embedding layer, and cross-entropy loss was used for training. |
| 139 | We observed that this improved the performance of extracted utterance embeddings for downstream tasks. |
| 140 | |
| 141 | The system is trained with recordings sampled at 16kHz (single channel). |
| 142 | The code will automatically normalize your audio (i.e., resampling + mono channel selection) when calling *classify_file* if needed. |
| 143 | |
| 144 | The model can classify a speech utterance according to the language spoken. |
| 145 | It covers 107 different languages ( |
| 146 | Abkhazian, |
| 147 | Afrikaans, |
| 148 | Amharic, |
| 149 | Arabic, |
| 150 | Assamese, |
| 151 | Azerbaijani, |
| 152 | Bashkir, |
| 153 | Belarusian, |
| 154 | Bulgarian, |
| 155 | Bengali, |
| 156 | Tibetan, |
| 157 | Breton, |
| 158 | Bosnian, |
| 159 | Catalan, |
| 160 | Cebuano, |
| 161 | Czech, |
| 162 | Welsh, |
| 163 | Danish, |
| 164 | German, |
| 165 | Greek, |
| 166 | English, |
| 167 | Esperanto, |
| 168 | Spanish, |
| 169 | Estonian, |
| 170 | Basque, |
| 171 | Persian, |
| 172 | Finnish, |
| 173 | Faroese, |
| 174 | French, |
| 175 | Galician, |
| 176 | Guarani, |
| 177 | Gujarati, |
| 178 | Manx, |
| 179 | Hausa, |
| 180 | Hawaiian, |
| 181 | Hindi, |
| 182 | Croatian, |
| 183 | Haitian, |
| 184 | Hungarian, |
| 185 | Armenian, |
| 186 | Interlingua, |
| 187 | Indonesian, |
| 188 | Icelandic, |
| 189 | Italian, |
| 190 | Hebrew, |
| 191 | Japanese, |
| 192 | Javanese, |
| 193 | Georgian, |
| 194 | Kazakh, |
| 195 | Central Khmer, |
| 196 | Kannada, |
| 197 | Korean, |
| 198 | Latin, |
| 199 | Luxembourgish, |
| 200 | Lingala, |
| 201 | Lao, |
| 202 | Lithuanian, |
| 203 | Latvian, |
| 204 | Malagasy, |
| 205 | Maori, |
| 206 | Macedonian, |
| 207 | Malayalam, |
| 208 | Mongolian, |
| 209 | Marathi, |
| 210 | Malay, |
| 211 | Maltese, |
| 212 | Burmese, |
| 213 | Nepali, |
| 214 | Dutch, |
| 215 | Norwegian Nynorsk, |
| 216 | Norwegian, |
| 217 | Occitan, |
| 218 | Panjabi, |
| 219 | Polish, |
| 220 | Pushto, |
| 221 | Portuguese, |
| 222 | Romanian, |
| 223 | Russian, |
| 224 | Sanskrit, |
| 225 | Scots, |
| 226 | Sindhi, |
| 227 | Sinhala, |
| 228 | Slovak, |
| 229 | Slovenian, |
| 230 | Shona, |
| 231 | Somali, |
| 232 | Albanian, |
| 233 | Serbian, |
| 234 | Sundanese, |
| 235 | Swedish, |
| 236 | Swahili, |
| 237 | Tamil, |
| 238 | Telugu, |
| 239 | Tajik, |
| 240 | Thai, |
| 241 | Turkmen, |
| 242 | Tagalog, |
| 243 | Turkish, |
| 244 | Tatar, |
| 245 | Ukrainian, |
| 246 | Urdu, |
| 247 | Uzbek, |
| 248 | Vietnamese, |
| 249 | Waray, |
| 250 | Yiddish, |
| 251 | Yoruba, |
| 252 | Mandarin Chinese). |
| 253 | |
| 254 | ## Intended uses & limitations |
| 255 | |
| 256 | The model has two uses: |
| 257 | |
| 258 | - use 'as is' for spoken language recognition |
| 259 | - use as an utterance-level feature (embedding) extractor, for creating a dedicated language ID model on your own data |
| 260 | |
| 261 | The model is trained on automatically collected YouTube data. For more |
| 262 | information about the dataset, see [here](https://cs.taltech.ee/staff/tanel.alumae/data/voxlingua107/). |
| 263 | |
| 264 | |
| 265 | #### How to use |
| 266 | ```bash |
| 267 | pip install git+https://github.com/speechbrain/speechbrain.git@develop |
| 268 | ``` |
| 269 | |
| 270 | ```python |
| 271 | import torchaudio |
| 272 | from speechbrain.inference.classifiers import EncoderClassifier |
| 273 | language_id = EncoderClassifier.from_hparams(source="speechbrain/lang-id-voxlingua107-ecapa", savedir="tmp") |
| 274 | # Download Thai language sample from Omniglot and cvert to suitable form |
| 275 | signal = language_id.load_audio("speechbrain/lang-id-voxlingua107-ecapa/udhr_th.wav") |
| 276 | prediction = language_id.classify_batch(signal) |
| 277 | print(prediction) |
| 278 | # (tensor([[-2.8646e+01, -3.0346e+01, -2.0748e+01, -2.9562e+01, -2.2187e+01, |
| 279 | # -3.2668e+01, -3.6677e+01, -3.3573e+01, -3.2545e+01, -2.4365e+01, |
| 280 | # -2.4688e+01, -3.1171e+01, -2.7743e+01, -2.9918e+01, -2.4770e+01, |
| 281 | # -3.2250e+01, -2.4727e+01, -2.6087e+01, -2.1870e+01, -3.2821e+01, |
| 282 | # -2.2128e+01, -2.2822e+01, -3.0888e+01, -3.3564e+01, -2.9906e+01, |
| 283 | # -2.2392e+01, -2.5573e+01, -2.6443e+01, -3.2429e+01, -3.2652e+01, |
| 284 | # -3.0030e+01, -2.4607e+01, -2.2967e+01, -2.4396e+01, -2.8578e+01, |
| 285 | # -2.5153e+01, -2.8475e+01, -2.6409e+01, -2.5230e+01, -2.7957e+01, |
| 286 | # -2.6298e+01, -2.3609e+01, -2.5863e+01, -2.8225e+01, -2.7225e+01, |
| 287 | # -3.0486e+01, -2.1185e+01, -2.7938e+01, -3.3155e+01, -1.9076e+01, |
| 288 | # -2.9181e+01, -2.2160e+01, -1.8352e+01, -2.5866e+01, -3.3636e+01, |
| 289 | # -4.2016e+00, -3.1581e+01, -3.1894e+01, -2.7834e+01, -2.5429e+01, |
| 290 | # -3.2235e+01, -3.2280e+01, -2.8786e+01, -2.3366e+01, -2.6047e+01, |
| 291 | # -2.2075e+01, -2.3770e+01, -2.2518e+01, -2.8101e+01, -2.5745e+01, |
| 292 | # -2.6441e+01, -2.9822e+01, -2.7109e+01, -3.0225e+01, -2.4566e+01, |
| 293 | # -2.9268e+01, -2.7651e+01, -3.4221e+01, -2.9026e+01, -2.6009e+01, |
| 294 | # -3.1968e+01, -3.1747e+01, -2.8156e+01, -2.9025e+01, -2.7756e+01, |
| 295 | # -2.8052e+01, -2.9341e+01, -2.8806e+01, -2.1636e+01, -2.3992e+01, |
| 296 | # -2.3794e+01, -3.3743e+01, -2.8332e+01, -2.7465e+01, -1.5085e-02, |
| 297 | # -2.9094e+01, -2.1444e+01, -2.9780e+01, -3.6046e+01, -3.7401e+01, |
| 298 | # -3.0888e+01, -3.3172e+01, -1.8931e+01, -2.2679e+01, -3.0225e+01, |
| 299 | # -2.4995e+01, -2.1028e+01]]), tensor([-0.0151]), tensor([94]), ['th']) |
| 300 | # The scores in the prediction[0] tensor can be interpreted as log-likelihoods that |
| 301 | # the given utterance belongs to the given language (i.e., the larger the better) |
| 302 | # The linear-scale likelihood can be retrieved using the following: |
| 303 | print(prediction[1].exp()) |
| 304 | # tensor([0.9850]) |
| 305 | # The identified language ISO code is given in prediction[3] |
| 306 | print(prediction[3]) |
| 307 | # ['th: Thai'] |
| 308 | |
| 309 | # Alternatively, use the utterance embedding extractor: |
| 310 | emb = language_id.encode_batch(signal) |
| 311 | print(emb.shape) |
| 312 | # torch.Size([1, 1, 256]) |
| 313 | ``` |
| 314 | To perform inference on the GPU, add `run_opts={"device":"cuda"}` when calling the `from_hparams` method. |
| 315 | |
| 316 | The system is trained with recordings sampled at 16kHz (single channel). |
| 317 | The code will automatically normalize your audio (i.e., resampling + mono channel selection) when calling *classify_file* if needed. Make sure your input tensor is compliant with the expected sampling rate if you use *encode_batch* and *classify_batch*. |
| 318 | |
| 319 | Warning: In the dataset and in the defaults of this model (see [`label_encoder.txt`](label_encoder.txt), the used ISO language code for Hebrew is obsolete (should be `he` instead of `iw`). The ISO language code for Javanese is incorrect (should be `jv` instead of `jw`). See [issue #2396](https://github.com/speechbrain/speechbrain/issues/2396). |
| 320 | |
| 321 | #### Limitations and bias |
| 322 | |
| 323 | Since the model is trained on VoxLingua107, it has many limitations and biases, some of which are: |
| 324 | |
| 325 | - Probably it's accuracy on smaller languages is quite limited |
| 326 | - Probably it works worse on female speech than male speech (because YouTube data includes much more male speech) |
| 327 | - Based on subjective experiments, it doesn't work well on speech with a foreign accent |
| 328 | - Probably it doesn't work well on children's speech and on persons with speech disorders |
| 329 | |
| 330 | |
| 331 | ## Training data |
| 332 | |
| 333 | The model is trained on [VoxLingua107](https://cs.taltech.ee/staff/tanel.alumae/data/voxlingua107/). |
| 334 | |
| 335 | VoxLingua107 is a speech dataset for training spoken language identification models. |
| 336 | The dataset consists of short speech segments automatically extracted from YouTube videos and labeled according the language of the video title and description, with some post-processing steps to filter out false positives. |
| 337 | |
| 338 | VoxLingua107 contains data for 107 languages. The total amount of speech in the training set is 6628 hours. |
| 339 | The average amount of data per language is 62 hours. However, the real amount per language varies a lot. There is also a seperate development set containing 1609 speech segments from 33 languages, validated by at least two volunteers to really contain the given language. |
| 340 | |
| 341 | ## Training procedure |
| 342 | |
| 343 | See the [SpeechBrain recipe](https://github.com/speechbrain/speechbrain/tree/voxlingua107/recipes/VoxLingua107/lang_id). |
| 344 | |
| 345 | ## Evaluation results |
| 346 | |
| 347 | Error rate: 6.7% on the VoxLingua107 development dataset |
| 348 | |
| 349 | #### Referencing SpeechBrain |
| 350 | ```bibtex |
| 351 | @misc{speechbrain, |
| 352 | title={{SpeechBrain}: A General-Purpose Speech Toolkit}, |
| 353 | author={Mirco Ravanelli and Titouan Parcollet and Peter Plantinga and Aku Rouhe and Samuele Cornell and Loren Lugosch and Cem Subakan and Nauman Dawalatabad and Abdelwahab Heba and Jianyuan Zhong and Ju-Chieh Chou and Sung-Lin Yeh and Szu-Wei Fu and Chien-Feng Liao and Elena Rastorgueva and François Grondin and William Aris and Hwidong Na and Yan Gao and Renato De Mori and Yoshua Bengio}, |
| 354 | year={2021}, |
| 355 | eprint={2106.04624}, |
| 356 | archivePrefix={arXiv}, |
| 357 | primaryClass={eess.AS}, |
| 358 | note={arXiv:2106.04624} |
| 359 | } |
| 360 | ``` |
| 361 | |
| 362 | ### Referencing VoxLingua107 |
| 363 | |
| 364 | ```bibtex |
| 365 | @inproceedings{valk2021slt, |
| 366 | title={{VoxLingua107}: a Dataset for Spoken Language Recognition}, |
| 367 | author={J{\"o}rgen Valk and Tanel Alum{\"a}e}, |
| 368 | booktitle={Proc. IEEE SLT Workshop}, |
| 369 | year={2021}, |
| 370 | } |
| 371 | ``` |
| 372 | |
| 373 | #### About SpeechBrain |
| 374 | SpeechBrain is an open-source and all-in-one speech toolkit. It is designed to be simple, extremely flexible, and user-friendly. Competitive or state-of-the-art performance is obtained in various domains. |
| 375 | Website: https://speechbrain.github.io/ |
| 376 | GitHub: https://github.com/speechbrain/speechbrain |
| 377 | |