File size: 7,803 Bytes
8b72146 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 |
---
language: rw
datasets:
- common_voice
metrics:
- wer
tags:
- audio
- automatic-speech-recognition
- speech
- xlsr-fine-tuning-week
license: apache-2.0
model-index:
- name: XLSR Wav2Vec2 Large Kinyarwanda with apostrophes
results:
- task:
name: Speech Recognition
type: automatic-speech-recognition
dataset:
name: Common Voice rw
type: common_voice
args: rw
metrics:
- name: Test WER
type: wer
value: 39.92
---
# Wav2Vec2-Large-XLSR-53-rw
Fine-tuned [facebook/wav2vec2-large-xlsr-53](https://huggingface.co/facebook/wav2vec2-large-xlsr-53) on Kinyarwanda using the [Common Voice](https://huggingface.co/datasets/common_voice) dataset, using about 25% of the training data (limited to utterances without downvotes and shorter with 9.5 seconds), and validated on 2048 utterances from the validation set. In contrast to the [lucio/wav2vec2-large-xlsr-kinyarwanda](https://huggingface.co/lucio/wav2vec2-large-xlsr-kinyarwanda) model, which does not predict any punctuation, this model attempts to predict the apostrophes that mark contractions of pronouns with vowel-initial words, but may overgeneralize.
When using this model, make sure that your speech input is sampled at 16kHz.
## 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
# WARNING! This will download and extract to use about 80GB on disk.
test_dataset = load_dataset("common_voice", "rw", split="test[:2%]")
processor = Wav2Vec2Processor.from_pretrained("lucio/wav2vec2-large-xlsr-kinyarwanda")
model = Wav2Vec2ForCTC.from_pretrained("lucio/wav2vec2-large-xlsr-kinyarwanda")
resampler = torchaudio.transforms.Resample(48_000, 16_000)
# Preprocessing the datasets.
# We need to read the audio 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[:2]["speech"], 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])
```
Result:
```
Prediction: ['yaherukago gukora igitaramo yiki mujyiwa na mor mu bubiligi', "ibi rero ntibizashoboka kandi n'umudabizi"]
Reference: ['Yaherukaga gukora igitaramo nk’iki mu Mujyi wa Namur mu Bubiligi.', 'Ibi rero, ntibizashoboka, kandi nawe arabizi.']
```
## Evaluation
The model can be evaluated as follows on the Kinyarwanda test data of Common Voice. Note that to even load the test data, the whole 40GB Kinyarwanda dataset will be downloaded and extracted into another 40GB directory, so you will need that space available on disk (e.g. not possible in the free tier of Google Colab). This script uses the `chunked_wer` function from [pcuenq](https://huggingface.co/pcuenq/wav2vec2-large-xlsr-53-es).
```python
import jiwer
import torch
import torchaudio
from datasets import load_dataset, load_metric
from transformers import Wav2Vec2ForCTC, Wav2Vec2Processor
import re
import unidecode
test_dataset = load_dataset("common_voice", "rw", split="test")
wer = load_metric("wer")
processor = Wav2Vec2Processor.from_pretrained("lucio/wav2vec2-large-xlsr-kinyarwanda-apostrophied")
model = Wav2Vec2ForCTC.from_pretrained("lucio/wav2vec2-large-xlsr-kinyarwanda-apostrophied")
model.to("cuda")
chars_to_ignore_regex = r'[!"#$%&()*+,./:;<=>?@\[\]\\_{}|~£¤¨©ª«¬®¯°·¸»¼½¾ðʺ˜˝ˮ‐–—―‚“”„‟•…″‽₋€™−√�]'
def remove_special_characters(batch):
batch["text"] = re.sub(r'[ʻʽʼ‘’´`]', r"'", batch["sentence"]) # normalize apostrophes
batch["text"] = re.sub(chars_to_ignore_regex, "", batch["text"]).lower().strip() # remove all other punctuation
batch["text"] = re.sub(r"([b-df-hj-np-tv-z])' ([aeiou])", r"\1'\2", batch["text"]) # remove spaces where apostrophe marks a deleted vowel
batch["text"] = re.sub(r"(-| '|' | +)", " ", batch["text"]) # treat dash and other apostrophes as word boundary
batch["text"] = unidecode.unidecode(batch["text"]) # strip accents from loanwords
return batch
## Audio pre-processing
resampler = torchaudio.transforms.Resample(48_000, 16_000)
def speech_file_to_array_fn(batch):
speech_array, sampling_rate = torchaudio.load(batch["path"])
batch["speech"] = resampler(speech_array).squeeze().numpy()
batch["sampling_rate"] = 16_000
return batch
def cv_prepare(batch):
batch = remove_special_characters(batch)
batch = speech_file_to_array_fn(batch)
return batch
test_dataset = test_dataset.map(cv_prepare)
# Preprocessing the datasets.
# We need to read the audio 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)
def chunked_wer(targets, predictions, chunk_size=None):
if chunk_size is None: return jiwer.wer(targets, predictions)
start = 0
end = chunk_size
H, S, D, I = 0, 0, 0, 0
while start < len(targets):
chunk_metrics = jiwer.compute_measures(targets[start:end], predictions[start:end])
H = H + chunk_metrics["hits"]
S = S + chunk_metrics["substitutions"]
D = D + chunk_metrics["deletions"]
I = I + chunk_metrics["insertions"]
start += chunk_size
end += chunk_size
return float(S + D + I) / float(H + S + D)
print("WER: {:2f}".format(100 * chunked_wer(result["sentence"], result["pred_strings"], chunk_size=4000)))
```
**Test Result**: 39.92 %
## Training
Examples from the Common Voice training dataset were used for training, after filtering out utterances that had any `down_vote` or were longer than 9.5 seconds. The data used totals about 125k examples, 25% of the available data, trained on 1 V100 GPU provided by OVHcloud, for a total of about 60 hours: 20 epochs on one block of 32k examples and then 10 epochs each on 3 more blocks of 32k examples. For validation, 2048 examples of the validation dataset were used.
The [script used for training](https://github.com/serapio/transformers/blob/feature/xlsr-finetune/examples/research_projects/wav2vec2/run_common_voice.py) is adapted from the [example script provided in the transformers repo](https://github.com/huggingface/transformers/blob/master/examples/research_projects/wav2vec2/run_common_voice.py). |