patrickvonplaten commited on
Commit
bf41557
β€’
1 Parent(s): 78d9d91

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +22 -1
README.md CHANGED
@@ -25,49 +25,67 @@ model-index:
25
  value: 29.95
26
  ---
27
  # Wav2Vec2-Large-XLSR-Latvian
 
28
  Fine-tuned [facebook/wav2vec2-large-xlsr-53](https://huggingface.co/facebook/wav2vec2-large-xlsr-53)
29
  on the [Latvian Common Voice dataset](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", "lv", split="test[:2%]")
 
39
  processor = Wav2Vec2Processor.from_pretrained("jimregan/wav2vec2-large-xlsr-latvian-cv")
40
  model = Wav2Vec2ForCTC.from_pretrained("jimregan/wav2vec2-large-xlsr-latvian-cv")
 
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 Latvian 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", "lv", split="test")
65
  wer = load_metric("wer")
 
66
  processor = Wav2Vec2Processor.from_pretrained("jimregan/wav2vec2-large-xlsr-latvian-cv")
67
  model = Wav2Vec2ForCTC.from_pretrained("jimregan/wav2vec2-large-xlsr-latvian-cv")
68
  model.to("cuda")
 
69
  chars_to_ignore_regex = '[,\?\.\!\;\:\"\β€œ\%\β€˜\”\(\)\*\…\β€”\–\']'
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):
@@ -75,7 +93,9 @@ def speech_file_to_array_fn(batch):
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 audio files as arrays
81
  def evaluate(batch):
@@ -85,8 +105,9 @@ def evaluate(batch):
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**: 29.95 %
92
- ```
 
25
  value: 29.95
26
  ---
27
  # Wav2Vec2-Large-XLSR-Latvian
28
+
29
  Fine-tuned [facebook/wav2vec2-large-xlsr-53](https://huggingface.co/facebook/wav2vec2-large-xlsr-53)
30
  on the [Latvian Common Voice dataset](https://huggingface.co/datasets/common_voice).
31
+
32
  When using this model, make sure that your speech input is sampled at 16kHz.
33
+
34
  ## Usage
35
  The model can be used directly (without a language model) as follows:
36
+
37
  ```python
38
  import torch
39
  import torchaudio
40
  from datasets import load_dataset
41
  from transformers import Wav2Vec2ForCTC, Wav2Vec2Processor
42
+
43
  test_dataset = load_dataset("common_voice", "lv", split="test[:2%]")
44
+
45
  processor = Wav2Vec2Processor.from_pretrained("jimregan/wav2vec2-large-xlsr-latvian-cv")
46
  model = Wav2Vec2ForCTC.from_pretrained("jimregan/wav2vec2-large-xlsr-latvian-cv")
47
+
48
  resampler = torchaudio.transforms.Resample(48_000, 16_000)
49
+
50
  # Preprocessing the datasets.
51
  # We need to read the aduio files as arrays
52
  def speech_file_to_array_fn(batch):
53
  speech_array, sampling_rate = torchaudio.load(batch["path"])
54
  batch["speech"] = resampler(speech_array).squeeze().numpy()
55
  return batch
56
+
57
  test_dataset = test_dataset.map(speech_file_to_array_fn)
58
  inputs = processor(test_dataset["speech"][:2], sampling_rate=16_000, return_tensors="pt", padding=True)
59
+
60
  with torch.no_grad():
61
  logits = model(inputs.input_values, attention_mask=inputs.attention_mask).logits
62
  predicted_ids = torch.argmax(logits, dim=-1)
63
+
64
  print("Prediction:", processor.batch_decode(predicted_ids))
65
  print("Reference:", test_dataset["sentence"][:2])
66
  ```
67
+
68
  ## Evaluation
69
+
70
  The model can be evaluated as follows on the Latvian test data of Common Voice.
71
+
72
  ```python
73
  import torch
74
  import torchaudio
75
  from datasets import load_dataset, load_metric
76
  from transformers import Wav2Vec2ForCTC, Wav2Vec2Processor
77
  import re
78
+
79
  test_dataset = load_dataset("common_voice", "lv", split="test")
80
  wer = load_metric("wer")
81
+
82
  processor = Wav2Vec2Processor.from_pretrained("jimregan/wav2vec2-large-xlsr-latvian-cv")
83
  model = Wav2Vec2ForCTC.from_pretrained("jimregan/wav2vec2-large-xlsr-latvian-cv")
84
  model.to("cuda")
85
+
86
  chars_to_ignore_regex = '[,\?\.\!\;\:\"\β€œ\%\β€˜\”\(\)\*\…\β€”\–\']'
87
  resampler = torchaudio.transforms.Resample(48_000, 16_000)
88
+
89
  # Preprocessing the datasets.
90
  # We need to read the aduio files as arrays
91
  def speech_file_to_array_fn(batch):
 
93
  speech_array, sampling_rate = torchaudio.load(batch["path"])
94
  batch["speech"] = resampler(speech_array).squeeze().numpy()
95
  return batch
96
+
97
  test_dataset = test_dataset.map(speech_file_to_array_fn)
98
+
99
  # Preprocessing the datasets.
100
  # We need to read the audio files as arrays
101
  def evaluate(batch):
 
105
  pred_ids = torch.argmax(logits, dim=-1)
106
  batch["pred_strings"] = processor.batch_decode(pred_ids)
107
  return batch
108
+
109
  result = test_dataset.map(evaluate, batched=True, batch_size=8)
110
  print("WER: {:2f}".format(100 * wer.compute(predictions=result["pred_strings"], references=result["sentence"])))
111
  ```
112
+
113
  **Test Result**: 29.95 %