Commit
•
95ce0ea
1
Parent(s):
c85b633
Create README.md
Browse files
README.md
ADDED
@@ -0,0 +1,109 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
---
|
2 |
+
language: en
|
3 |
+
datasets:
|
4 |
+
- librispeech_asr
|
5 |
+
tags:
|
6 |
+
- speech
|
7 |
+
- audio
|
8 |
+
- automatic-speech-recognition
|
9 |
+
- hf-asr-leaderboard
|
10 |
+
license: apache-2.0
|
11 |
+
model-index:
|
12 |
+
- name: wav2vec2-conformer-rel-pos-large-960h-ft
|
13 |
+
results:
|
14 |
+
- task:
|
15 |
+
name: Automatic Speech Recognition
|
16 |
+
type: automatic-speech-recognition
|
17 |
+
dataset:
|
18 |
+
name: Librispeech (clean)
|
19 |
+
type: librispeech_asr
|
20 |
+
args: en
|
21 |
+
metrics:
|
22 |
+
- name: Test WER
|
23 |
+
type: wer
|
24 |
+
value: 1.85
|
25 |
+
---
|
26 |
+
|
27 |
+
# Wav2Vec2-Conformer-Large-960h with Rotary Position Embeddings
|
28 |
+
|
29 |
+
[Facebook's Wav2Vec2 Conformer (TODO-add link)]()
|
30 |
+
|
31 |
+
Wav2Vec2 Conformer with rotary position embeddings, pretrained and fine-tuned on 960 hours of Librispeech on 16kHz sampled speech audio. When using the model make sure that your speech input is also sampled at 16Khz.
|
32 |
+
|
33 |
+
[Paper (TODO)](https://arxiv.org/abs/2006.11477)
|
34 |
+
|
35 |
+
Authors: ...
|
36 |
+
|
37 |
+
**Abstract**
|
38 |
+
|
39 |
+
...
|
40 |
+
|
41 |
+
The original model can be found under https://github.com/pytorch/fairseq/tree/master/examples/wav2vec#wav2vec-20.
|
42 |
+
|
43 |
+
|
44 |
+
# Usage
|
45 |
+
|
46 |
+
To transcribe audio files the model can be used as a standalone acoustic model as follows:
|
47 |
+
|
48 |
+
```python
|
49 |
+
from transformers import Wav2Vec2Processor, Wav2Vec2ConformerForCTC
|
50 |
+
from datasets import load_dataset
|
51 |
+
import torch
|
52 |
+
|
53 |
+
# load model and processor
|
54 |
+
processor = Wav2Vec2Processor.from_pretrained("facebook/wav2vec2-conformer-rope-large-960h-ft")
|
55 |
+
model = Wav2Vec2ConformerForCTC.from_pretrained("facebook/wav2vec2-conformer-rope-large-960h-ft")
|
56 |
+
|
57 |
+
# load dummy dataset and read soundfiles
|
58 |
+
ds = load_dataset("patrickvonplaten/librispeech_asr_dummy", "clean", split="validation")
|
59 |
+
|
60 |
+
# tokenize
|
61 |
+
input_values = processor(ds[0]["audio"]["array"], return_tensors="pt", padding="longest").input_values
|
62 |
+
|
63 |
+
# retrieve logits
|
64 |
+
logits = model(input_values).logits
|
65 |
+
|
66 |
+
# take argmax and decode
|
67 |
+
predicted_ids = torch.argmax(logits, dim=-1)
|
68 |
+
transcription = processor.batch_decode(predicted_ids)
|
69 |
+
```
|
70 |
+
|
71 |
+
## Evaluation
|
72 |
+
|
73 |
+
This code snippet shows how to evaluate **facebook/wav2vec2-conformer-rope-large-960h-ft** on LibriSpeech's "clean" and "other" test data.
|
74 |
+
|
75 |
+
```python
|
76 |
+
from datasets import load_dataset
|
77 |
+
from transformers import Wav2Vec2ConformerForCTC, Wav2Vec2Processor
|
78 |
+
import torch
|
79 |
+
from jiwer import wer
|
80 |
+
|
81 |
+
|
82 |
+
librispeech_eval = load_dataset("librispeech_asr", "clean", split="test")
|
83 |
+
|
84 |
+
model = Wav2Vec2ConformerForCTC.from_pretrained("facebook/wav2vec2-conformer-rope-large-960h-ft").to("cuda")
|
85 |
+
processor = Wav2Vec2Processor.from_pretrained("facebook/wav2vec2-conformer-rope-large-960h-ft")
|
86 |
+
|
87 |
+
def map_to_pred(batch):
|
88 |
+
inputs = processor(batch["audio"]["array"], return_tensors="pt", padding="longest")
|
89 |
+
input_values = inputs.input_values.to("cuda")
|
90 |
+
attention_mask = inputs.attention_mask.to("cuda")
|
91 |
+
|
92 |
+
with torch.no_grad():
|
93 |
+
logits = model(input_values, attention_mask=attention_mask).logits
|
94 |
+
|
95 |
+
predicted_ids = torch.argmax(logits, dim=-1)
|
96 |
+
transcription = processor.batch_decode(predicted_ids)
|
97 |
+
batch["transcription"] = transcription
|
98 |
+
return batch
|
99 |
+
|
100 |
+
result = librispeech_eval.map(map_to_pred, remove_columns=["audio"])
|
101 |
+
|
102 |
+
print("WER:", wer(result["text"], result["transcription"]))
|
103 |
+
```
|
104 |
+
|
105 |
+
*Result (WER)*:
|
106 |
+
|
107 |
+
| "clean" | "other" |
|
108 |
+
|---|---|
|
109 |
+
| 1.96 | 3.98 |
|