Hugging Face's logo

badrex
/
mms-300m-upper-sorbian

Automatic Speech Recognition
Transformers
Safetensors
PyTorch
Upper Sorbian
wav2vec2
Upper-Sorbian
MMS
Eval Results
Inference Endpoints
Model card Files Community
mms-300m-upper-sorbian / README.md
badrex's picture
badrex
Update README.md
482c9b6 verified
preview code
|
raw
history blame contribute delete
4.55 kB
metadata
license: cc-by-nc-4.0
base_model: facebook/mms-300m
language: hsb
tags:
  - automatic-speech-recognition
  - Upper-Sorbian
  - pytorch
  - transformers
  - MMS
datasets:
  - common_voice_17_0
metrics:
  - wer
model-index:
  - name: mms-300m-upper-sorbian
    results:
      - task:
          name: Automatic Speech Recognition
          type: automatic-speech-recognition
        dataset:
          name: common_voice_17_0
          type: common_voice_17_0
          config: hsb
          split: validation
          args: hsb
        metrics:
          - name: Wer
            type: wer
            value: 0.42025316455696204
pipeline_tag: automatic-speech-recognition
library_name: transformers

Visualize in Weights & Biases

mms-300m-upper-sorbian

This is an automatic speech recognition (ASR) model for the Upper Sorbian language, a minority Slavic language spoken in Saxony, Germany. The model is a fine-tuned version of facebook/mms-300m and trained on the train split of Common Voice 17 dataset (Upper Sorbian - hsb).

It achieves the following results on the evaluation set (validation split):

  • Loss: 0.6600
  • Wer: 0.4203
  • Cer: 0.0930

Model description

ASR model trained on crowdsourced speech from Mozilla Common Voice. It can be used to transcribe Upper Sorbian speech into text.

Intended uses & limitations

The model is intended to be used as a speech-to-text system. However, it has only been trained on scripted read speech thus it may not perform well on conversational speech.

Training and evaluation data

Mozilla Common Voice (Upper Sorbian - hsb)

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

  • learning_rate: 0.0003
  • train_batch_size: 16
  • eval_batch_size: 8
  • seed: 42
  • gradient_accumulation_steps: 2
  • total_train_batch_size: 32
  • optimizer: Adam with betas=(0.9,0.98) and epsilon=1e-08
  • lr_scheduler_type: linear
  • lr_scheduler_warmup_ratio: 0.1
  • num_epochs: 100
  • mixed_precision_training: Native AMP

Training results

Training Loss Epoch Step Validation Loss Wer Cer
3.408 3.9216 100 3.3797 1.0 1.0
3.1402 7.8431 200 3.1629 1.0 1.0
0.7479 11.7647 300 1.0200 0.9323 0.2916
0.2111 15.6863 400 0.7733 0.7095 0.1844
0.1842 19.6078 500 0.7090 0.6051 0.1549
0.0618 23.5294 600 0.7410 0.6184 0.1474
0.0802 27.4510 700 0.7037 0.55 0.1308
0.0392 31.3725 800 0.7951 0.5924 0.1430
0.0504 35.2941 900 0.7686 0.5418 0.1290
0.0436 39.2157 1000 0.7336 0.55 0.1239
0.0282 43.1373 1100 0.7303 0.5133 0.1211
0.0333 47.0588 1200 0.6966 0.5057 0.1204
0.0243 50.9804 1300 0.6883 0.4734 0.1088
0.0218 54.9020 1400 0.7155 0.5051 0.1168
0.0219 58.8235 1500 0.6778 0.4943 0.1111
0.0101 62.7451 1600 0.6565 0.4570 0.1063
0.012 66.6667 1700 0.6723 0.4405 0.1016
0.0233 70.5882 1800 0.6700 0.4589 0.1039
0.0075 74.5098 1900 0.7376 0.4570 0.1062
0.0165 78.4314 2000 0.7359 0.4443 0.1010
0.0071 82.3529 2100 0.7349 0.4532 0.1022
0.0055 86.2745 2200 0.6797 0.4411 0.0991
0.0051 90.1961 2300 0.7313 0.4354 0.0975
0.0062 94.1176 2400 0.6847 0.4203 0.0938
0.0142 98.0392 2500 0.6600 0.4203 0.0930

Framework versions

  • Transformers 4.42.0.dev0
  • Pytorch 2.3.1+cu121
  • Datasets 2.19.2
  • Tokenizers 0.19.1