---
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