README.md
| 1 | --- |
| 2 | library_name: transformers |
| 3 | license: apache-2.0 |
| 4 | base_model: facebook/wav2vec2-xls-r-300m |
| 5 | tags: |
| 6 | - generated_from_trainer |
| 7 | - speech-recognition |
| 8 | - audio-classification |
| 9 | - voicemail-detection |
| 10 | model-index: |
| 11 | - name: wav2vec-vm-finetune |
| 12 | results: [] |
| 13 | language: |
| 14 | - en |
| 15 | metrics: |
| 16 | - accuracy |
| 17 | --- |
| 18 | |
| 19 | # wav2vec-vm-finetune |
| 20 | |
| 21 | This model is a fine-tuned version of [facebook/wav2vec2-xls-r-300m](https://huggingface.co/facebook/wav2vec2-xls-r-300m) for **voicemail detection**. It is trained on a dataset of call recordings to distinguish between **voicemail greetings** and **live human responses**. |
| 22 | |
| 23 | ## Model description |
| 24 | |
| 25 | This model builds on **wav2vec2-xls-r-300m**, a self-supervised speech model trained on large-scale multilingual data. We fine-tuned it on the first two seconds of a call. |
| 26 | |
| 27 | ## Intended uses & limitations |
| 28 | |
| 29 | - Automated voicemail detection in AI-powered call assistants. |
| 30 | - Filtering voicemail responses in customer service and sales call automation. |
| 31 | |
| 32 | - Only trianed on the English language. |
| 33 | - Assumes the voicemail track is isolated and contains no audio from the caller. |
| 34 | - Designed for the first two seconds of audio when calling a voicemail. |
| 35 | |
| 36 | ## Training and evaluation data |
| 37 | |
| 38 | The model was trained on a proprietary dataset of call recordings, labeled as: |
| 39 | - **Live human responses** |
| 40 | - **Voicemail greetings** |
| 41 | |
| 42 | The dataset includes diverse voicemail recordings across multiple types to improve generalization. |
| 43 | |
| 44 | |
| 45 | ## Evaluation metrics |
| 46 | |
| 47 | The model achieved: |
| 48 | - **98% accuracy** on voicemail detection. |
| 49 | |
| 50 | |
| 51 | ## Training procedure |
| 52 | |
| 53 | ### Training hyperparameters |
| 54 | |
| 55 | The following hyperparameters were used during training: |
| 56 | - learning_rate: 0.0003 |
| 57 | - train_batch_size: 16 |
| 58 | - eval_batch_size: 8 |
| 59 | - seed: 42 |
| 60 | - gradient_accumulation_steps: 2 |
| 61 | - total_train_batch_size: 32 |
| 62 | - optimizer: Use OptimizerNames.ADAMW_TORCH with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments |
| 63 | - lr_scheduler_type: linear |
| 64 | - lr_scheduler_warmup_steps: 500 |
| 65 | - num_epochs: 10 |
| 66 | - mixed_precision_training: Native AMP |
| 67 | |
| 68 | ### Framework versions |
| 69 | |
| 70 | - Transformers 4.48.2 |
| 71 | - Pytorch 2.5.1+cu124 |
| 72 | - Datasets 1.18.3 |
| 73 | - Tokenizers 0.21.0 |