README.md
| 1 | --- |
| 2 | language: te |
| 3 | datasets: |
| 4 | - openslr |
| 5 | metrics: |
| 6 | - wer |
| 7 | tags: |
| 8 | - audio |
| 9 | - automatic-speech-recognition |
| 10 | - speech |
| 11 | - xlsr-fine-tuning-week |
| 12 | license: apache-2.0 |
| 13 | model-index: |
| 14 | - name: Anurag Singh XLSR Wav2Vec2 Large 53 Telugu |
| 15 | results: |
| 16 | - task: |
| 17 | name: Speech Recognition |
| 18 | type: automatic-speech-recognition |
| 19 | dataset: |
| 20 | name: OpenSLR te |
| 21 | type: openslr |
| 22 | args: te |
| 23 | metrics: |
| 24 | - name: Test WER |
| 25 | type: wer |
| 26 | value: 44.98 |
| 27 | --- |
| 28 | # Wav2Vec2-Large-XLSR-53-Telugu |
| 29 | Fine-tuned [facebook/wav2vec2-large-xlsr-53](https://huggingface.co/facebook/wav2vec2-large-xlsr-53) on Telugu using the [OpenSLR SLR66](http://openslr.org/66/) dataset. |
| 30 | When using this model, make sure that your speech input is sampled at 16kHz. |
| 31 | ## Usage |
| 32 | The model can be used directly (without a language model) as follows: |
| 33 | ```python |
| 34 | import torch |
| 35 | import torchaudio |
| 36 | from datasets import load_dataset |
| 37 | from transformers import Wav2Vec2ForCTC, Wav2Vec2Processor |
| 38 | import pandas as pd |
| 39 | # Evaluation notebook contains the procedure to download the data |
| 40 | df = pd.read_csv("/content/te/test.tsv", sep="\t") |
| 41 | df["path"] = "/content/te/clips/" + df["path"] |
| 42 | test_dataset = Dataset.from_pandas(df) |
| 43 | processor = Wav2Vec2Processor.from_pretrained("anuragshas/wav2vec2-large-xlsr-53-telugu") |
| 44 | model = Wav2Vec2ForCTC.from_pretrained("anuragshas/wav2vec2-large-xlsr-53-telugu") |
| 45 | resampler = torchaudio.transforms.Resample(48_000, 16_000) |
| 46 | # Preprocessing the datasets. |
| 47 | # We need to read the aduio files as arrays |
| 48 | def speech_file_to_array_fn(batch): |
| 49 | speech_array, sampling_rate = torchaudio.load(batch["path"]) |
| 50 | batch["speech"] = resampler(speech_array).squeeze().numpy() |
| 51 | return batch |
| 52 | test_dataset = test_dataset.map(speech_file_to_array_fn) |
| 53 | inputs = processor(test_dataset["speech"][:2], sampling_rate=16_000, return_tensors="pt", padding=True) |
| 54 | with torch.no_grad(): |
| 55 | logits = model(inputs.input_values, attention_mask=inputs.attention_mask).logits |
| 56 | predicted_ids = torch.argmax(logits, dim=-1) |
| 57 | print("Prediction:", processor.batch_decode(predicted_ids)) |
| 58 | print("Reference:", test_dataset["sentence"][:2]) |
| 59 | ``` |
| 60 | ## Evaluation |
| 61 | ```python |
| 62 | import torch |
| 63 | import torchaudio |
| 64 | from datasets import Dataset, load_metric |
| 65 | from transformers import Wav2Vec2ForCTC, Wav2Vec2Processor |
| 66 | import re |
| 67 | from sklearn.model_selection import train_test_split |
| 68 | import pandas as pd |
| 69 | # Evaluation notebook contains the procedure to download the data |
| 70 | df = pd.read_csv("/content/te/test.tsv", sep="\t") |
| 71 | df["path"] = "/content/te/clips/" + df["path"] |
| 72 | test_dataset = Dataset.from_pandas(df) |
| 73 | wer = load_metric("wer") |
| 74 | processor = Wav2Vec2Processor.from_pretrained("anuragshas/wav2vec2-large-xlsr-53-telugu") |
| 75 | model = Wav2Vec2ForCTC.from_pretrained("anuragshas/wav2vec2-large-xlsr-53-telugu") |
| 76 | model.to("cuda") |
| 77 | chars_to_ignore_regex = '[\,\?\.\!\-\_\;\:\"\“\%\‘\”\।\’\'\&]' |
| 78 | resampler = torchaudio.transforms.Resample(48_000, 16_000) |
| 79 | def normalizer(text): |
| 80 | # Use your custom normalizer |
| 81 | text = text.replace("\\n","\n") |
| 82 | text = ' '.join(text.split()) |
| 83 | text = re.sub(r'''([a-z]+)''','',text,flags=re.IGNORECASE) |
| 84 | text = re.sub(r'''%'''," శాతం ", text) |
| 85 | text = re.sub(r'''(/|-|_)'''," ", text) |
| 86 | text = re.sub("ై","ై", text) |
| 87 | text = text.strip() |
| 88 | return text |
| 89 | def speech_file_to_array_fn(batch): |
| 90 | batch["sentence"] = normalizer(batch["sentence"]) |
| 91 | batch["sentence"] = re.sub(chars_to_ignore_regex, '', batch["sentence"]).lower()+ " " |
| 92 | speech_array, sampling_rate = torchaudio.load(batch["path"]) |
| 93 | batch["speech"] = resampler(speech_array).squeeze().numpy() |
| 94 | return batch |
| 95 | test_dataset = test_dataset.map(speech_file_to_array_fn) |
| 96 | # Preprocessing the datasets. |
| 97 | # We need to read the aduio files as arrays |
| 98 | def evaluate(batch): |
| 99 | inputs = processor(batch["speech"], sampling_rate=16_000, return_tensors="pt", padding=True) |
| 100 | with torch.no_grad(): |
| 101 | logits = model(inputs.input_values.to("cuda"), attention_mask=inputs.attention_mask.to("cuda")).logits |
| 102 | pred_ids = torch.argmax(logits, dim=-1) |
| 103 | batch["pred_strings"] = processor.batch_decode(pred_ids) |
| 104 | return batch |
| 105 | result = test_dataset.map(evaluate, batched=True, batch_size=8) |
| 106 | print("WER: {:2f}".format(100 * wer.compute(predictions=result["pred_strings"], references=result["sentence"]))) |
| 107 | ``` |
| 108 | |
| 109 | **Test Result**: 44.98% |
| 110 | ## Training |
| 111 | 70% of the OpenSLR Telugu dataset was used for training. |
| 112 | |
| 113 | Train Split of annotations is [here](https://www.dropbox.com/s/xqc0wtour7f9h4c/train.tsv) |
| 114 | |
| 115 | Test Split of annotations is [here](https://www.dropbox.com/s/qw1uy63oj4qdiu4/test.tsv) |
| 116 | |
| 117 | Training Data Preparation notebook can be found [here](https://colab.research.google.com/drive/1_VR1QtY9qoiabyXBdJcOI29-xIKGdIzU?usp=sharing) |
| 118 | |
| 119 | Training notebook can be found[here](https://colab.research.google.com/drive/14N-j4m0Ng_oktPEBN5wiUhDDbyrKYt8I?usp=sharing) |
| 120 | |
| 121 | Evaluation notebook is [here](https://colab.research.google.com/drive/1SLEvbTWBwecIRTNqpQ0fFTqmr1-7MnSI?usp=sharing) |