deepdml

This repository contains the conversion of distil-whisper/distil-large-v3.5 to the CTranslate2 model format. This model can be used in CTranslate2 or projects based on CTranslate2 such as faster-whisper. Note that the model weights are saved in FP16. This type can be changed when the model is loaded using the `computetype` option in CTranslate2. For more information about the original model, see its model card.

This model is a fine-tuned version of deepdml/whisper-large-v3-turbo on the google/fleurs dataset. It achieves the following results on the evaluation set: - Loss: 0.7028 - Wer: 31.2646 - Cer: 10.8084 The following hyperparameters were used during training: - learningrate: 1e-05 - trainbatchsize: 16 - evalbatchsize: 16 - seed: 42 - optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08 - lrschedulertype: linear - lrschedulerwarmupratio: 0.04 - trainingsteps: 5000 | Training Loss | Epoch | Step | Validation Loss | Wer | Cer | |:-------------:|:-----:|:----:|:---------------:|:-------:|:-------:| | 0.2019 | 0.2 | 1000 | 0.6438 | 36.4596 | 12.5223 | | 0.1293 | 0.4 | 2000 | 0.6558 | 33.7633 | 11.6044 | | 0.0589 | 0.6 | 3000 | 0.6882 | 31.8758 | 10.6653 | | 0.0504 | 0.8 | 4000 | 0.6845 | 31.0669 | 10.3172 | | 0.0353 | 1.0 | 5000 | 0.7028 | 31.2646 | 10.8084 | - Transformers 4.42.0.dev0 - Pytorch 2.3.0+cu121 - Datasets 2.19.1 - Tokenizers 0.19.1 Please cite the model using the following BibTeX entry:

This model is a fine-tuned version of openai/whisper-medium on the Common Voice 17.0 dataset. It achieves the following results on the evaluation set: - Loss: 2.4504 - Wer: 15.3250 The following hyperparameters were used during training: - learningrate: 1e-05 - trainbatchsize: 32 - evalbatchsize: 8 - seed: 42 - optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08 - lrschedulertype: linear - lrschedulerwarmupsteps: 500 - trainingsteps: 5000 | Training Loss | Epoch | Step | Validation Loss | Wer | |:-------------:|:-----:|:----:|:---------------:|:-------:| | 0.2423 | 0.2 | 1000 | 2.0205 | 21.3178 | | 0.0667 | 0.4 | 2000 | 2.3750 | 18.2033 | | 0.047 | 0.6 | 3000 | 2.4276 | 17.5658 | | 0.0249 | 0.8 | 4000 | 2.7231 | 16.1576 | | 0.017 | 1.0 | 5000 | 2.4504 | 15.3250 | - Transformers 4.42.0.dev0 - Pytorch 2.3.0+cu121 - Datasets 2.19.1 - Tokenizers 0.19.1

This model is a fine-tuned version of openai/whisper-small on the Common Voice 17.0 dataset. It achieves the following results on the evaluation set: - Loss: 2.2980 - Wer: 26.5362 The following hyperparameters were used during training: - learningrate: 1e-05 - trainbatchsize: 64 - evalbatchsize: 8 - seed: 42 - optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08 - lrschedulertype: linear - lrschedulerwarmupsteps: 500 - trainingsteps: 5000 | Training Loss | Epoch | Step | Validation Loss | Wer | |:-------------:|:-----:|:----:|:---------------:|:-------:| | 0.2564 | 0.2 | 1000 | 1.8598 | 30.7045 | | 0.061 | 0.4 | 2000 | 2.1891 | 28.2575 | | 0.0368 | 0.6 | 3000 | 2.2045 | 26.3524 | | 0.0262 | 0.8 | 4000 | 2.4023 | 26.1855 | | 0.0128 | 1.0 | 5000 | 2.2980 | 26.5362 | - Transformers 4.42.0.dev0 - Pytorch 2.3.0+cu121 - Datasets 2.19.1 - Tokenizers 0.19.1

This model is a fine-tuned version of openai/whisper-base on the google/fleurs dataset. It achieves the following results on the evaluation set: - Loss: 1.0933 - Wer: 54.9487 - Cer: 21.3532 The following hyperparameters were used during training: - learningrate: 1e-05 - trainbatchsize: 64 - evalbatchsize: 64 - seed: 42 - optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08 - lrschedulertype: linear - lrschedulerwarmupratio: 0.04 - trainingsteps: 5000 | Training Loss | Epoch | Step | Validation Loss | Wer | Cer | |:-------------:|:------:|:----:|:---------------:|:-------:|:-------:| | 0.2087 | 0.2 | 1000 | 0.8427 | 54.4143 | 20.1160 | | 0.0734 | 1.0814 | 2000 | 0.9702 | 55.5707 | 21.6200 | | 0.0609 | 1.2814 | 3000 | 1.0272 | 54.0256 | 20.4927 | | 0.0336 | 2.1628 | 4000 | 1.0804 | 54.4337 | 20.4677 | | 0.0341 | 3.0442 | 5000 | 1.0933 | 54.9487 | 21.3532 | - Transformers 4.42.0.dev0 - Pytorch 2.3.0+cu121 - Datasets 2.19.1 - Tokenizers 0.19.1 Please cite the model using the following BibTeX entry:

This model is a fine-tuned version of deepdml/whisper-large-v3-turbo on the Common Voice 17.0 dataset. It achieves the following results on the evaluation set: - Loss: 1.0072 - Wer: 13.1125 The following hyperparameters were used during training: - learningrate: 1e-05 - trainbatchsize: 8 - evalbatchsize: 8 - seed: 42 - optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08 - lrschedulertype: linear - lrschedulerwarmupratio: 0.04 - trainingsteps: 15000 | Training Loss | Epoch | Step | Validation Loss | Wer | |:-------------:|:-----:|:-----:|:---------------:|:-------:| | 0.0084 | 1.0 | 15000 | 1.0072 | 13.1125 | - Transformers 4.42.0.dev0 - Pytorch 2.3.0+cu121 - Datasets 2.19.1 - Tokenizers 0.19.1

This model is a fine-tuned version of openai/whisper-small on the Quran dataset. It achieves the following results on the evaluation set: - Loss: 1.1161 - Wer: 11.9246 The following hyperparameters were used during training: - learningrate: 1e-05 - trainbatchsize: 32 - evalbatchsize: 32 - seed: 42 - optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08 - lrschedulertype: linear - lrschedulerwarmupratio: 0.04 - trainingsteps: 41811 | Training Loss | Epoch | Step | Validation Loss | Wer | |:-------------:|:------:|:-----:|:---------------:|:-------:| | 0.069 | 0.0239 | 1000 | 0.6098 | 18.3529 | | 0.0331 | 0.0478 | 2000 | 0.6068 | 21.3135 | | 0.0362 | 0.0718 | 3000 | 0.6658 | 12.6054 | | 0.0119 | 0.0957 | 4000 | 0.7669 | 77.1810 | | 0.0097 | 0.1196 | 5000 | 0.7483 | 0.7036 | | 0.0465 | 0.1435 | 6000 | 0.8107 | 71.6345 | | 0.0103 | 0.1674 | 7000 | 0.8574 | 0.5642 | | 0.01 | 0.1913 | 8000 | 0.9754 | 66.6750 | | 0.0088 | 0.2153 | 9000 | 0.8795 | 6.2410 | | 0.006 | 0.2392 | 10000 | 0.9043 | 29.0623 | | 0.007 | 0.2631 | 11000 | 1.0010 | 82.3004 | | 0.0056 | 0.2870 | 12000 | 0.9681 | 59.1593 | | 0.0103 | 0.3109 | 13000 | 1.0676 | 52.7744 | | 0.0157 | 0.3348 | 14000 | 1.0371 | 37.3935 | | 0.0036 | 0.3588 | 15000 | 0.9451 | 0.1759 | | 0.0108 | 0.3827 | 16000 | 0.9329 | 0.4752 | | 0.0042 | 0.4066 | 17000 | 0.9960 | 45.4826 | | 0.0052 | 0.4305 | 18000 | 1.0721 | 77.3752 | | 0.004 | 0.4544 | 19000 | 0.9658 | 0.1850 | | 0.0167 | 1.0072 | 20000 | 0.9981 | 11.4495 | | 0.0046 | 1.0311 | 21000 | 1.0030 | 68.6625 | | 0.0072 | 1.0550 | 22000 | 0.9918 | 34.6865 | | 0.0084 | 1.0789 | 23000 | 1.0646 | 90.7870 | | 0.0046 | 1.1028 | 24000 | 1.1421 | 82.4877 | | 0.0076 | 1.1267 | 25000 | 1.0337 | 1.9966 | | 0.0046 | 1.1507 | 26000 | 1.0411 | 5.1742 | | 0.0046 | 1.1746 | 27000 | 1.0758 | 1.0988 | | 0.0057 | 1.1985 | 28000 | 1.0331 | 18.5471 | | 0.0012 | 1.2224 | 29000 | 1.0715 | 0.2901 | | 0.0015 | 1.2463 | 30000 | 1.0870 | 0.1005 | | 0.0032 | 1.2702 | 31000 | 1.0910 | 17.0029 | | 0.0016 | 1.2942 | 32000 | 1.1208 | 48.9275 | | 0.0021 | 1.3181 | 33000 | 1.1441 | 0.0366 | | 0.0022 | 1.3420 | 34000 | 1.2570 | 76.4797 | | 0.0124 | 1.3659 | 35000 | 1.1071 | 3.5043 | | 0.0106 | 1.3898 | 36000 | 1.1721 | 42.3507 | | 0.011 | 1.4137 | 37000 | 1.1471 | 4.5688 | | 0.0107 | 1.4377 | 38000 | 1.1831 | 26.8624 | | 0.0034 | 1.4616 | 39000 | 1.1529 | 7.2302 | | 0.0111 | 2.0148 | 40000 | 1.1303 | 42.4854 | | 0.0077 | 2.0387 | 41000 | 1.1161 | 11.9246 | - Transformers 4.42.0.dev0 - Pytorch 2.3.0+cu121 - Datasets 2.19.1 - Tokenizers 0.19.1 Please cite the model using the following BibTeX entry:

This model is a fine-tuned version of openai/whisper-base on the Quran dataset. It achieves the following results on the evaluation set: - Loss: 0.0129 - Wer: 16.1651 The following hyperparameters were used during training: - learningrate: 1e-05 - trainbatchsize: 64 - evalbatchsize: 64 - seed: 42 - optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08 - lrschedulertype: linear - lrschedulerwarmupratio: 0.04 - trainingsteps: 20905 | Training Loss | Epoch | Step | Validation Loss | Wer | |:-------------:|:------:|:-----:|:---------------:|:-------:| | 0.1147 | 0.0478 | 1000 | 0.1044 | 46.2933 | | 0.0414 | 0.0957 | 2000 | 0.0608 | 56.1421 | | 0.0888 | 0.1435 | 3000 | 0.0417 | 38.5954 | | 0.0238 | 0.1913 | 4000 | 0.0383 | 24.4833 | | 0.0171 | 0.2392 | 5000 | 0.0216 | 17.2703 | | 0.0166 | 0.2870 | 6000 | 0.0228 | 24.1871 | | 0.0243 | 0.3348 | 7000 | 0.0302 | 40.0101 | | 0.0307 | 0.3827 | 8000 | 0.0200 | 23.0494 | | 0.0081 | 0.4305 | 9000 | 0.0198 | 13.9295 | | 0.01 | 1.0071 | 10000 | 0.0226 | 25.6162 | | 0.0169 | 1.0550 | 11000 | 0.0146 | 13.7364 | | 0.0148 | 1.1028 | 12000 | 0.0189 | 32.5028 | | 0.0137 | 1.1506 | 13000 | 0.0308 | 68.5616 | | 0.012 | 1.1985 | 14000 | 0.0176 | 17.1361 | | 0.0116 | 1.2463 | 15000 | 0.0128 | 17.6105 | | 0.0098 | 1.2941 | 16000 | 0.0221 | 31.6383 | | 0.0124 | 1.3420 | 17000 | 0.0180 | 26.3024 | | 0.0067 | 1.3898 | 18000 | 0.0188 | 36.3544 | | 0.0166 | 1.4376 | 19000 | 0.0138 | 26.3494 | | 0.0154 | 2.0148 | 20000 | 0.0129 | 16.1651 | - Transformers 4.42.0.dev0 - Pytorch 2.3.0+cu121 - Datasets 2.19.1 - Tokenizers 0.19.1 Please cite the model using the following BibTeX entry:

This model is a fine-tuned version of openai/whisper-small on the google/fleurs dataset. It achieves the following results on the evaluation set: - Loss: 1.5879 - Wer: 46.1037 The following hyperparameters were used during training: - learningrate: 1e-05 - trainbatchsize: 64 - evalbatchsize: 8 - seed: 42 - optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08 - lrschedulertype: linear - lrschedulerwarmupsteps: 500 - trainingsteps: 5000 | Training Loss | Epoch | Step | Validation Loss | Wer | |:-------------:|:------:|:----:|:---------------:|:-------:| | 0.1171 | 0.2 | 1000 | 1.2732 | 44.9937 | | 0.028 | 1.0814 | 2000 | 1.4495 | 46.2251 | | 0.0277 | 1.2814 | 3000 | 1.4894 | 45.3892 | | 0.0084 | 2.1628 | 4000 | 1.5629 | 44.6881 | | 0.0065 | 3.0442 | 5000 | 1.5879 | 46.1037 | - Transformers 4.42.0.dev0 - Pytorch 2.3.0+cu121 - Datasets 2.19.1 - Tokenizers 0.19.1

This model is a fine-tuned version of openai/whisper-tiny on the Quran dataset. It achieves the following results on the evaluation set: - Loss: 0.0188 - Wer: 1.9035 - Cer: 0.6582 The following hyperparameters were used during training: - learningrate: 1e-05 - trainbatchsize: 128 - evalbatchsize: 128 - seed: 42 - optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08 - lrschedulertype: linear - lrschedulerwarmupratio: 0.04 - trainingsteps: 10452 | Training Loss | Epoch | Step | Validation Loss | Wer | Cer | |:-------------:|:------:|:-----:|:---------------:|:-------:|:------:| | 0.1204 | 0.0957 | 1000 | 0.1922 | 19.2088 | 6.2785 | | 0.0939 | 0.1914 | 2000 | 0.0947 | 9.7640 | 3.1603 | | 0.0229 | 0.2870 | 3000 | 0.0615 | 6.4970 | 2.1582 | | 0.0345 | 0.3827 | 4000 | 0.0447 | 4.6754 | 1.5399 | | 0.0222 | 1.0071 | 5000 | 0.0353 | 3.6451 | 1.2228 | | 0.0165 | 1.1028 | 6000 | 0.0302 | 3.2147 | 1.0847 | | 0.0148 | 1.1984 | 7000 | 0.0252 | 2.6315 | 0.9019 | | 0.006 | 1.2941 | 8000 | 0.0222 | 2.3478 | 0.7940 | | 0.0121 | 1.3898 | 9000 | 0.0200 | 2.0813 | 0.7232 | | 0.0078 | 2.0147 | 10000 | 0.0188 | 1.9035 | 0.6582 | - Transformers 4.42.0.dev0 - Pytorch 2.3.0+cu121 - Datasets 2.19.1 - Tokenizers 0.19.1 Please cite the model using the following BibTeX entry:

This model is a fine-tuned version of openai/whisper-base on the Common Voice 17.0 dataset. It achieves the following results on the evaluation set: - Loss: 2.5272 - Wer: 62.6842 The following hyperparameters were used during training: - learningrate: 1e-05 - trainbatchsize: 64 - evalbatchsize: 8 - seed: 42 - optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08 - lrschedulertype: linear - lrschedulerwarmupsteps: 500 - trainingsteps: 5000 | Training Loss | Epoch | Step | Validation Loss | Wer | |:-------------:|:------:|:----:|:---------------:|:-------:| | 0.3483 | 0.2 | 1000 | 2.0647 | 67.7943 | | 0.1912 | 1.0454 | 2000 | 2.3245 | 65.8907 | | 0.131 | 1.2454 | 3000 | 2.4512 | 63.3511 | | 0.0954 | 2.0908 | 4000 | 2.4555 | 62.8998 | | 0.0711 | 2.2908 | 5000 | 2.5272 | 62.6842 | - Transformers 4.42.0.dev0 - Pytorch 2.3.0+cu121 - Datasets 2.19.1 - Tokenizers 0.19.1

This model is a fine-tuned version of openai/whisper-medium on the Common Voice 17.0 dataset. It achieves the following results on the evaluation set: - Loss: 1.0757 - Wer: 0.1553 The following hyperparameters were used during training: - learningrate: 1e-05 - trainbatchsize: 16 - evalbatchsize: 8 - seed: 42 - optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08 - lrschedulertype: linear - lrschedulerwarmupratio: 0.04 - trainingsteps: 15000 | Training Loss | Epoch | Step | Validation Loss | Wer | |:-------------:|:-----:|:-----:|:---------------:|:------:| | 0.0377 | 1.0 | 15000 | 1.0757 | 0.1553 | - Transformers 4.42.0.dev0 - Pytorch 2.3.0+cu121 - Datasets 2.19.1 - Tokenizers 0.19.1