wav2vec2-jv-base-openslr

This model is a fine-tuned version of facebook/wav2vec2-base on the OpenSLR41 datasets. It achieves the following results on the evaluation set:

• Loss: 0.2843
• Wer: 0.1502

Model description

The model is a fine-tuned version of wav2vec2, specifically adapted using the OpenSLR 41 dataset, which is focused on the Javanese language domain. This adaptation enables the model to effectively recognize and process spoken Javanese, leveraging the robust capabilities of the wav2vec2 architecture combined with domain-specific training data.

Intended uses & limitations

This model is intended for transcribing spoken Javanese language from audio recordings. It achieves a Word Error Rate (WER) of 15%, indicating that while the model performs reasonably well, it still produces significant transcription errors. Users should be aware that the accuracy may vary, particularly in cases with challenging audio conditions or less common dialects. Additionally, this model requires input audio at a sample rate of 16kHz, which may limit its applicability for recordings at different sample rates or lower quality audio files.

Training and evaluation data

The model use OpenSLR41 datasets, and split into 2 section (training and testing), then the model is trained using 1xA100 GPU with a training duration of 4-5 hours.

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 0.0001
• train_batch_size: 8
• eval_batch_size: 8
• seed: 42
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 1000
• num_epochs: 65
• mixed_precision_training: Native AMP

Log Data | Training results

Training Loss	Epoch	Step	Validation Loss	Wer
0.5361	2.8329	2000	0.4626	0.4238
0.332	5.6657	4000	0.3857	0.3749
0.242	8.4986	6000	0.3456	0.3060
0.1893	11.3314	8000	0.3250	0.2846
0.1566	14.1643	10000	0.3260	0.2640
0.1433	16.9972	12000	0.2891	0.2516
0.124	19.8300	14000	0.3172	0.2433
0.1103	22.6629	16000	0.3099	0.2453
0.1015	25.4958	18000	0.3087	0.2295
0.088	28.3286	20000	0.3250	0.2054
0.0831	31.1615	22000	0.3127	0.2143
0.0748	33.9943	24000	0.2973	0.1923
0.0696	36.8272	26000	0.3103	0.2026
0.0622	39.6601	28000	0.3292	0.2068
0.0564	42.4929	30000	0.2965	0.1916
0.0507	45.3258	32000	0.3061	0.1819
0.0475	48.1586	34000	0.2784	0.1881
0.0448	50.9915	36000	0.2872	0.1764
0.0413	53.8244	38000	0.2854	0.1716
0.0357	56.6572	40000	0.2862	0.1723
0.0328	59.4901	42000	0.2887	0.1654
0.0324	62.3229	44000	0.2843	0.1502

How to run (Gradio Web)

import torch
import torchaudio
import gradio as gr
import numpy as np
from transformers import pipeline, AutoProcessor, AutoModelForCTC

device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

# Load the model and processor
MODEL_NAME = "<fill this to your model>"
processor = AutoProcessor.from_pretrained(MODEL_NAME)
model = AutoModelForCTC.from_pretrained(MODEL_NAME)

# Move model to GPU
model.to(device)

# Create the pipeline with the model and processor
transcriber = pipeline("automatic-speech-recognition", model=model, tokenizer=processor.tokenizer, feature_extractor=processor.feature_extractor, device=device)

def transcribe(audio):
    sr, y = audio
    y = y.astype(np.float32)
    y /= np.max(np.abs(y))

    return transcriber({"sampling_rate": sr, "raw": y})["text"]

demo = gr.Interface(
    transcribe,
    gr.Audio(sources=["upload"]),
    "text",
)

demo.launch(share=True)

How to run

import torch
import torchaudio
import gradio as gr
import numpy as np
from transformers import pipeline, AutoProcessor, AutoModelForCTC

device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

# Load the model and processor
MODEL_NAME = "<fill this to actual model>"
processor = AutoProcessor.from_pretrained(MODEL_NAME)
model = AutoModelForCTC.from_pretrained(MODEL_NAME)

# Move model to GPU
model.to(device)

# Load audio file
AUDIO_PATH = "<replace 'path_to_audio_file.wav' with the actual path to your audio file>"
audio_input, sample_rate = torchaudio.load(AUDIO_PATH)

# Ensure the audio is mono (1 channel)
if audio_input.shape[0] > 1:
    audio_input = torch.mean(audio_input, dim=0, keepdim=True)

# Resample audio if necessary
if sample_rate != 16000:
    resampler = torchaudio.transforms.Resample(orig_freq=sample_rate, new_freq=16000)
    audio_input = resampler(audio_input)

# Process the audio input
input_values = processor(audio_input.squeeze(), sampling_rate=16000, return_tensors="pt").input_values

# Move input values to GPU
input_values = input_values.to(device)

# Perform inference
with torch.no_grad():
    logits = model(input_values).logits

# Decode the logits to text
predicted_ids = torch.argmax(logits, dim=-1)
transcription = processor.batch_decode(predicted_ids)[0]

print("Transcription:", transcription)

Framework versions

• Transformers 4.44.0
• Pytorch 2.2.1+cu118
• Datasets 2.20.0
• Tokenizers 0.19.1