1 ---
2 language: sah
3 datasets:
4 - common_voice
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 Sakha
15 results:
16 - task:
17 name: Speech Recognition
18 type: automatic-speech-recognition
19 dataset:
20 name: Common Voice sah
21 type: common_voice
22 args: sah
23 metrics:
24 - name: Test WER
25 type: wer
26 value: 38.04
27 ---
28 # Wav2Vec2-Large-XLSR-53-Sakha
29 Fine-tuned [facebook/wav2vec2-large-xlsr-53](https://huggingface.co/facebook/wav2vec2-large-xlsr-53) on Sakha using the [Common Voice](https://huggingface.co/datasets/common_voice).
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 test_dataset = load_dataset("common_voice", "sah", split="test[:2%]")
39 processor = Wav2Vec2Processor.from_pretrained("anuragshas/wav2vec2-large-xlsr-53-sah")
40 model = Wav2Vec2ForCTC.from_pretrained("anuragshas/wav2vec2-large-xlsr-53-sah")
41 resampler = torchaudio.transforms.Resample(48_000, 16_000)
42 # Preprocessing the datasets.
43 # We need to read the aduio files as arrays
44 def speech_file_to_array_fn(batch):
45 speech_array, sampling_rate = torchaudio.load(batch["path"])
46 batch["speech"] = resampler(speech_array).squeeze().numpy()
47 return batch
48 test_dataset = test_dataset.map(speech_file_to_array_fn)
49 inputs = processor(test_dataset["speech"][:2], sampling_rate=16_000, return_tensors="pt", padding=True)
50 with torch.no_grad():
51 logits = model(inputs.input_values, attention_mask=inputs.attention_mask).logits
52 predicted_ids = torch.argmax(logits, dim=-1)
53 print("Prediction:", processor.batch_decode(predicted_ids))
54 print("Reference:", test_dataset["sentence"][:2])
55 ```
56 ## Evaluation
57 The model can be evaluated as follows on the Sakha test data of Common Voice.
58 ```python
59 import torch
60 import torchaudio
61 from datasets import load_dataset, load_metric
62 from transformers import Wav2Vec2ForCTC, Wav2Vec2Processor
63 import re
64 test_dataset = load_dataset("common_voice", "sah", split="test")
65 wer = load_metric("wer")
66 processor = Wav2Vec2Processor.from_pretrained("anuragshas/wav2vec2-large-xlsr-53-sah")
67 model = Wav2Vec2ForCTC.from_pretrained("anuragshas/wav2vec2-large-xlsr-53-sah")
68 model.to("cuda")
69 chars_to_ignore_regex = '[\,\?\.\!\-\;\:\"\“\%\”\„\–\…\«\»]'
70 resampler = torchaudio.transforms.Resample(48_000, 16_000)
71 # Preprocessing the datasets.
72 # We need to read the aduio files as arrays
73 def speech_file_to_array_fn(batch):
74 batch["sentence"] = re.sub(chars_to_ignore_regex, '', batch["sentence"]).lower()
75 speech_array, sampling_rate = torchaudio.load(batch["path"])
76 batch["speech"] = resampler(speech_array).squeeze().numpy()
77 return batch
78 test_dataset = test_dataset.map(speech_file_to_array_fn)
79 # Preprocessing the datasets.
80 # We need to read the aduio files as arrays
81 def evaluate(batch):
82 inputs = processor(batch["speech"], sampling_rate=16_000, return_tensors="pt", padding=True)
83 with torch.no_grad():
84 logits = model(inputs.input_values.to("cuda"), attention_mask=inputs.attention_mask.to("cuda")).logits
85 pred_ids = torch.argmax(logits, dim=-1)
86 batch["pred_strings"] = processor.batch_decode(pred_ids)
87 return batch
88 result = test_dataset.map(evaluate, batched=True, batch_size=8)
89 print("WER: {:2f}".format(100 * wer.compute(predictions=result["pred_strings"], references=result["sentence"])))
90 ```
91 **Test Result**: 38.04 %
92 ## Training
93 The Common Voice `train` and `validation` datasets were used for training.