File size: 6,402 Bytes
c5471f3 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 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 |
---
language:
- eo
license: apache-2.0
tags:
- automatic-speech-recognition
- mozilla-foundation/common_voice_13_0
- generated_from_trainer
datasets:
- common_voice_13_0
metrics:
- wer
- cer
model-index:
- name: wav2vec2-common_voice_13_0-eo-10
results:
- task:
name: Automatic Speech Recognition
type: automatic-speech-recognition
dataset:
name: mozilla-foundation/common_voice_13_0
type: common_voice_13_0
config: eo
split: validation
args: 'Config: eo, Training split: train, Eval split: validation'
metrics:
- name: WER
type: wer
value: 0.0656526475637132
- name: CER
type: cer
value: 0.0118
---
# wav2vec2-common_voice_13_0-eo-10, an Esperanto speech recognizer
This model is a fine-tuned version of [facebook/wav2vec2-large-xlsr-53](https://huggingface.co./facebook/wav2vec2-large-xlsr-53) on the [mozilla-foundation/common_voice_13_0](https://huggingface.co./datasets/mozilla-foundation/common_voice_13_0) Esperanto dataset.
It achieves the following results on the evaluation set:
- Loss: 0.0453
- Cer: 0.0118
- Wer: 0.0657
The first 10 examples in the evaluation set:
| Actual<br>Predicted | CER |
|:--------------------|:----|
| `la orienta parto apud benino kaj niĝerio estis nomita sklavmarbordo`<br>`la orienta parto apud benino kaj niĝerio estis nomita sklafmarbordo` | 0.014925373134328358 |
| `en la sekva jaro li ricevis premion`<br>`en la sekva jaro li ricevis premion` | 0.0 |
| `ŝi studis historion ĉe la universitato de brita kolumbio`<br>`ŝi studis historion ĉe la universitato de brita kolumbio` | 0.0 |
| `larĝaj ŝtupoj kuras al la fasado`<br>`larĝaj ŝtupoj kuras al la fasado` | 0.0 |
| `la municipo ĝuas duan epokon de etendo kaj disvolviĝo`<br>`la municipo ĝuas duan eepokon de etendo kaj disvolviĝo` | 0.018867924528301886 |
| `li estis ankaŭ katedrestro kaj dekano`<br>`li estis ankaŭ katedristo kaj dekano` | 0.05405405405405406 |
| `librovendejo apartenas al la muzeo`<br>`librovendejo apartenas al la muzeo` | 0.0 |
| `ĝi estas kutime malfacile videbla kaj troviĝas en subkreskaĵaro de arbaroj`<br>`ĝi estas kutime malfacile videbla kaj troviĝas en subkreskaĵo de arbaroj` | 0.02702702702702703 |
| `unue ili estas ruĝaj poste brunaj`<br>`unue ili estas ruĝaj poste brunaj` | 0.0 |
| `la loĝantaro laboras en la proksima ĉefurbo`<br>`la loĝantaro laboras en la proksima ĉefurbo` | 0.0 |
## Model description
See [facebook/wav2vec2-large-xlsr-53](https://huggingface.co./facebook/wav2vec2-large-xlsr-53).
## Intended uses & limitations
Speech recognition for Esperanto. The base model was pretrained and finetuned on 16kHz sampled speech audio. When using the model make sure that your speech input is also sampled at 16KHz.
The output is all lowercase, no punctuation.
## Training and evaluation data
The training split was set to `train` while the eval split was set to `validation`. Some files were filtered out of the train and validation dataset due to bad data; see [xekri/wav2vec2-common_voice_13_0-eo-3](https://huggingface.co./xekri/wav2vec2-common_voice_13_0-eo-3) for a detailed discussion. In summary, I used `xekri/wav2vec2-common_voice_13_0-eo-3` as a detector to detect bad files, then hardcoded those files into the trainer code to be filtered out.
## Training procedure
I used a modified version of [`run_speech_recognition_ctc.py`](https://github.com/huggingface/transformers/tree/main/examples/pytorch/speech-recognition) for training. See [`run_speech_recognition_ctc.py`](https://huggingface.co./xekri/wav2vec2-common_voice_13_0-eo-10/blob/main/run_speech_recognition_ctc.py) in this repo.
The parameters to the trainer are in [train.json](https://huggingface.co./xekri/wav2vec2-common_voice_13_0-eo-10/blob/main/train.json) in this repo.
The key changes between this training run and `xekri/wav2vec2-common_voice_13_0-eo-3`, aside from the filtering and use of the full training and validation sets are:
* Layer drop probability is 20%
* Train only for 5 epochs
### Training hyperparameters
The following hyperparameters were used during training:
- learning_rate: 3e-05
- train_batch_size: 16
- eval_batch_size: 8
- seed: 42
- gradient_accumulation_steps: 2
- total_train_batch_size: 32
- layerdrop: 0.2
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
- lr_scheduler_type: linear
- lr_scheduler_warmup_steps: 500
- num_epochs: 5
- mixed_precision_training: Native AMP
### Training results
| Training Loss | Epoch | Step | Cer | Validation Loss | Wer |
|:-------------:|:-----:|:-----:|:------:|:---------------:|:------:|
| 2.9894 | 0.22 | 1000 | 1.0 | 2.9257 | 1.0 |
| 0.7104 | 0.44 | 2000 | 0.0457 | 0.2129 | 0.2538 |
| 0.2853 | 0.67 | 3000 | 0.0274 | 0.1109 | 0.1583 |
| 0.2327 | 0.89 | 4000 | 0.0231 | 0.0909 | 0.1320 |
| 0.1917 | 1.11 | 5000 | 0.0206 | 0.0775 | 0.1188 |
| 0.1803 | 1.33 | 6000 | 0.0184 | 0.0698 | 0.1055 |
| 0.1661 | 1.56 | 7000 | 0.0169 | 0.0645 | 0.0961 |
| 0.1635 | 1.78 | 8000 | 0.0170 | 0.0639 | 0.0964 |
| 0.1555 | 2.0 | 9000 | 0.0156 | 0.0592 | 0.0881 |
| 0.1386 | 2.22 | 10000 | 0.0147 | 0.0559 | 0.0821 |
| 0.1338 | 2.45 | 11000 | 0.0146 | 0.0548 | 0.0831 |
| 0.1307 | 2.67 | 12000 | 0.0137 | 0.0529 | 0.0759 |
| 0.1297 | 2.89 | 13000 | 0.0134 | 0.0504 | 0.0745 |
| 0.1201 | 3.11 | 14000 | 0.0131 | 0.0499 | 0.0734 |
| 0.1152 | 3.34 | 15000 | 0.0128 | 0.0484 | 0.0712 |
| 0.1144 | 3.56 | 16000 | 0.0125 | 0.0477 | 0.0695 |
| 0.1179 | 3.78 | 17000 | 0.0122 | 0.0468 | 0.0679 |
| 0.1112 | 4.0 | 18000 | 0.0121 | 0.0468 | 0.0676 |
| 0.1141 | 4.23 | 19000 | 0.0121 | 0.0462 | 0.0668 |
| 0.1085 | 4.45 | 20000 | 0.0119 | 0.0458 | 0.0664 |
| 0.105 | 4.67 | 21000 | 0.0119 | 0.0456 | 0.0660 |
| 0.1072 | 4.89 | 22000 | 0.0119 | 0.0454 | 0.0658 |
### Framework versions
- Transformers 4.29.2
- Pytorch 2.0.1+cu117
- Datasets 2.12.0
- Tokenizers 0.13.3
|