wer 50.95, +less60wer.ipynb

README.md

@@ -23,4 +23,101 @@ model-index:
          value: 69.11
 ---
 
-First full try on the new dataset. Updates to come.
+# Wav2Vec2-Large-XLSR-53-ncj/nah
+
+Fine-tuned [facebook/wav2vec2-large-xlsr-53](https://huggingface.co/facebook/wav2vec2-large-xlsr-53) on Nahuatl specifically of the Nort of Puebla (ncj) using a derivate of [SLR92](https://www.openslr.org/92/), and some samples of `es` and `de` datasets from [Common Voice](https://huggingface.co/datasets/common_voice).
+
+## Usage
+
+The model can be used directly (without a language model) as follows:
+
+```python
+import torch
+import torchaudio
+from datasets import load_dataset
+from transformers import Wav2Vec2ForCTC, Wav2Vec2Processor
+
+test_dataset = load_dataset("common_voice", "{lang_id}", split="test[:2%]") # TODO: publish nahuatl_slr92_by_sentence
+
+processor = Wav2Vec2Processor.from_pretrained("tyoc213/wav2vec2-large-xlsr-nahuatl")
+model = Wav2Vec2ForCTC.from_pretrained("tyoc213/wav2vec2-large-xlsr-nahuatl")
+
+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])
+```
+
+
+## Evaluation
+
+The model can be evaluated as follows on the Nahuatl specifically of the Nort of Puebla (ncj) test data of Common Voice.
+
+
+```python
+import torch
+import torchaudio
+from datasets import load_dataset, load_metric
+from transformers import Wav2Vec2ForCTC, Wav2Vec2Processor
+import re
+
+test_dataset = load_dataset("common_voice", "{lang_id}", split="test") # TODO: publish nahuatl_slr92_by_sentence
+wer = load_metric("wer")
+
+processor = Wav2Vec2Processor.from_pretrained("tyoc213/wav2vec2-large-xlsr-nahuatl")
+model = Wav2Vec2ForCTC.from_pretrained("tyoc213/wav2vec2-large-xlsr-nahuatl")
+model.to("cuda")
+
+chars_to_ignore_regex = '[\,\?\.\!\-\;\"\“\%\‘\”\�\(\)\-]'
+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):
+	batch["sentence"] = re.sub(chars_to_ignore_regex, '', batch["sentence"]).lower()
+	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)
+
+# Preprocessing the datasets.
+# We need to read the aduio files as arrays
+def evaluate(batch):
+	inputs = processor(batch["speech"], sampling_rate=16_000, return_tensors="pt", padding=True)
+
+	with torch.no_grad():
+		logits = model(inputs.input_values.to("cuda"), attention_mask=inputs.attention_mask.to("cuda")).logits
+
+	pred_ids = torch.argmax(logits, dim=-1)
+	batch["pred_strings"] = processor.batch_decode(pred_ids)
+	return batch
+
+result = test_dataset.map(evaluate, batched=True, batch_size=8)
+
+print("WER: {:2f}".format(100 * wer.compute(predictions=result["pred_strings"], references=result["sentence"])))
+```
+
+**Test Result**: 50.95 %
+
+
+## Training
+
+A derivate of [SLR92](https://www.openslr.org/92/) to be published soon.And some samples of `es` and `de` datasets from [Common Voice](https://huggingface.co/datasets/common_voice)
+
+The script used for training can be found [less60wer.ipynb](./less60wer.ipynb)

config.json

@@ -70,7 +70,7 @@
   "num_conv_pos_embeddings": 128,
   "num_feat_extract_layers": 7,
   "num_hidden_layers": 24,
-  "pad_token_id": 42,
+  "pad_token_id": 44,
   "transformers_version": "4.5.0.dev0",
-  "vocab_size": 43
+  "vocab_size": 45
 }

pytorch_model.bin

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:41e78fb1cf5485a524c97d2b63d3669629a1a5a037282d4d810f3c294d402a29
-size 1262110103
+oid sha256:f05c8cadb48e7e8aa1177ab3193ad36b7871134416dbaccd5c773220cf44dcc2
+size 1262118359

vocab.json

@@ -1 +1 @@
-{"k": 0, "ú": 1, "¿": 2, "v": 3, "]": 4, "x": 5, "c": 6, "{": 7, "f": 8, "d": 9, "i": 10, "t": 11, "j": 12, "a": 13, "y": 14, "e": 15, "é": 16, "z": 17, "'": 18, "[": 19, "u": 20, "*": 21, "¡": 22, "r": 23, "ñ": 24, "q": 25, "á": 26, "s": 27, "b": 29, "´": 30, "m": 31, "ó": 32, "l": 33, "p": 34, "í": 35, "o": 36, "w": 37, "g": 38, "h": 39, "n": 40, "|": 28, "[UNK]": 41, "[PAD]": 42}
+{"x": 0, "v": 1, "]": 2, "í": 3, ":": 4, "k": 5, "y": 6, "ö": 7, "'": 8, "h": 9, "¿": 11, "ñ": 12, "n": 13, "ü": 14, "ä": 15, "t": 16, "m": 17, "s": 18, "g": 19, "á": 20, "z": 21, "o": 22, "w": 23, "[": 24, "r": 25, "b": 26, "ß": 27, "d": 28, "ó": 29, "i": 30, "e": 31, "„": 32, "ú": 33, "c": 34, "f": 35, "p": 36, "a": 37, "l": 38, "q": 39, "j": 40, "u": 41, "é": 42, "|": 10, "[UNK]": 43, "[PAD]": 44}