model_KWS

This model is a fine-tuned version of facebook/wav2vec2-base on the audiofolder dataset. It achieves the following results on the evaluation set:

• Loss: 0.3346
• Accuracy: 0.9825

Model description

Finetuned on custom commands: "ambient", "light", "off", "on", "scene1", "scene2", "scene3", "void"

Intended uses & limitations

Intended for keyword spotting applications.

Training and evaluation data

3200 training samples, 800 testing samples in total. Originally was recorded 20 samples of every class. Each sample was randomly augmented with random methods: pitch-shifting, time-stretching, volume-change, gaussian noise.

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 3e-05
• train_batch_size: 32
• eval_batch_size: 32
• seed: 42
• gradient_accumulation_steps: 4
• total_train_batch_size: 128
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_ratio: 0.1
• num_epochs: 10

Training results

Training Loss	Epoch	Step	Validation Loss	Accuracy
2.0119	1.0	25	1.9832	0.375
1.4505	2.0	50	1.3361	0.8337
1.0767	3.0	75	0.8700	0.955
0.7448	4.0	100	0.6919	0.9513
0.6143	5.0	125	0.5333	0.9625
0.4924	6.0	150	0.4387	0.98
0.4544	7.0	175	0.3844	0.985
0.3888	8.0	200	0.3668	0.9812
0.3734	9.0	225	0.3436	0.9825
0.3522	10.0	250	0.3346	0.9825

Framework versions

• Transformers 4.31.0
• Pytorch 2.0.1
• Datasets 2.14.0
• Tokenizers 0.13.3