Disclaimer: This is a work in progress. See VoxRex for more details.
Finetuned version of KBs VoxRex large model using Swedish radio broadcasts, NST and Common Voice data. Evalutation without a language model gives the following: WER for NST + Common Voice test set (2% of total sentences) is 3.617%. WER for Common Voice test set is 9.914% directly and 7.77% with a 4-gram language model.
When using this model, make sure that your speech input is sampled at 16kHz.
This model has been fine-tuned for 120000 updates on NST + CommonVoice and then for an additional 20000 updates on CommonVoice only. The additional fine-tuning on CommonVoice hurts performance on the NST+CommonVoice test set somewhat and, unsurprisingly, improves it on the CommonVoice test set. It seems to perform generally better though .
The model can be used directly (without a language model) as follows:
import torch import torchaudio from datasets import load_dataset from transformers import Wav2Vec2ForCTC, Wav2Vec2Processor test_dataset = load_dataset("common_voice", "sv-SE", split="test[:2%]"). processor = Wav2Vec2Processor.from_pretrained("KBLab/wav2vec2-large-voxrex-swedish") model = Wav2Vec2ForCTC.from_pretrained("KBLab/wav2vec2-large-voxrex-swedish") resampler = torchaudio.transforms.Resample(48_000, 16_000) # Preprocessing the datasets. # We need to read the aduio files as arrays def speech_file_to_array_fn(batch): speech_array, sampling_rate = torchaudio.load(batch["path"]) batch["speech"] = resampler(speech_array).squeeze().numpy() return batch test_dataset = test_dataset.map(speech_file_to_array_fn) inputs = processor(test_dataset["speech"][:2], sampling_rate=16_000, return_tensors="pt", padding=True) with torch.no_grad(): logits = model(inputs.input_values, attention_mask=inputs.attention_mask).logits predicted_ids = torch.argmax(logits, dim=-1) print("Prediction:", processor.batch_decode(predicted_ids)) print("Reference:", test_dataset["sentence"][:2])
- Downloads last month