Commit
·
1a99978
1
Parent(s):
db7fd0e
Update README.md
Browse files
README.md
CHANGED
@@ -4,4 +4,90 @@ task_categories:
|
|
4 |
- automatic-speech-recognition
|
5 |
language:
|
6 |
- bn
|
7 |
-
---
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
4 |
- automatic-speech-recognition
|
5 |
language:
|
6 |
- bn
|
7 |
+
---
|
8 |
+
|
9 |
+
# MegaBNSpeech
|
10 |
+
|
11 |
+
This model is based on a study aimed at tackling one of the primary challenges in developing Automatic Speech Recognition (ASR) for low-resource languages (Bangla): the limited access to domain-specific labeled data. To address this, the study introduces a pseudo-labeling approach to develop a domain-agnostic ASR dataset.
|
12 |
+
|
13 |
+
The methodology led to the creation of a robust 20k+ hours labeled Bangla speech dataset, which encompasses a wide variety of topics, speaking styles, dialects, noisy environments, and conversational scenarios. Using this data, a conformer-based ASR system was designed. The effectiveness of the model, especially when trained on pseudo-labeled data, was benchmarked against publicly available datasets and compared with other models. The research promises that experimental resources stemming from this study will be made publicly available.
|
14 |
+
|
15 |
+
## How to use:
|
16 |
+
|
17 |
+
The datasets library provides the capability to load and process your dataset efficiently using just Python. You can easily download and set up the dataset on your local drive with a single call using the *load_dataset* function.
|
18 |
+
```python
|
19 |
+
from datasets import load_dataset
|
20 |
+
dataset = load_dataset("hishab/MegaBNSpeech", split="train")
|
21 |
+
```
|
22 |
+
|
23 |
+
With the datasets library, you have the option to stream the dataset in real-time by appending the streaming=True parameter to the load_dataset function. In streaming mode, the dataset loads one sample at a time instead of storing the whole dataset on the disk.
|
24 |
+
```python
|
25 |
+
from datasets import load_dataset
|
26 |
+
dataset = load_dataset("hishab/MegaBNSpeech", split="train", streaming=True)
|
27 |
+
print(next(iter(dataset)))
|
28 |
+
```
|
29 |
+
## Speech Recognition (ASR)
|
30 |
+
|
31 |
+
```python
|
32 |
+
from datasets import load_dataset
|
33 |
+
|
34 |
+
mega_bn_asr = load_dataset("hishab/MegaBNSpeech")
|
35 |
+
|
36 |
+
# see structure
|
37 |
+
print(mega_bn_asr)
|
38 |
+
|
39 |
+
# load audio sample on the fly
|
40 |
+
audio_input = mega_bn_asr["train"][0]["audio"] # first decoded audio sample
|
41 |
+
transcription = mega_bn_asr["train"][0]["transcription"] # first transcription
|
42 |
+
# use `audio_input` and `transcription` to fine-tune your model for ASR
|
43 |
+
```
|
44 |
+
## Data Structure
|
45 |
+
- The dataset was developed using a pseudo-labeling approach.
|
46 |
+
- The largest collection of Bangla audio-video data was curated and cleaned from various Bangla TV channels on YouTube. This data covers varying domains, speaking styles, dialects, and communication channels.
|
47 |
+
- Alignments from two ASR systems were leveraged to segment and automatically annotate the audio segments.
|
48 |
+
- The created dataset was used to design an end-to-end state-of-the-art Bangla ASR system.
|
49 |
+
|
50 |
+
### Data Instances
|
51 |
+
- Size of downloaded dataset files: ___ GB
|
52 |
+
- Size of the generated dataset: ___ MB
|
53 |
+
- Total amount of disk used: ___ GB
|
54 |
+
|
55 |
+
An example of a data instance looks as follows:
|
56 |
+
```
|
57 |
+
{
|
58 |
+
"id": 0,
|
59 |
+
"channel_id": "UCPREnbhKQP-hsVfsfKP-mCw",
|
60 |
+
"channel_name": "NEWS24",
|
61 |
+
"video_id": "2kux6rFXMeM",
|
62 |
+
"audio_path": "data/train/wav/UCPREnbhKQP-hsVfsfKP-mCw_id_2kux6rFXMeM_85.wav",
|
63 |
+
"transcription": "পরীক্ষার মূল্য তালিকা উন্মুক্ত স্থানে প্রদর্শনের আদেশ দেন এই আদেশ পাওয়ার",
|
64 |
+
"duration": 5.055
|
65 |
+
}
|
66 |
+
```
|
67 |
+
### Data Fields
|
68 |
+
The data fields are written below.
|
69 |
+
- **id** (int): ID of audio sample
|
70 |
+
- **num_samples** (int): Number of float values
|
71 |
+
- **path** (str): Path to the audio file
|
72 |
+
- **audio** (dict): Audio object including loaded audio array, sampling rate and path ot audio
|
73 |
+
- **raw_transcription** (str): The non-normalized transcription of the audio file
|
74 |
+
- **transcription** (str): Transcription of the audio file
|
75 |
+
- **lang_id** (int): Class id of language
|
76 |
+
|
77 |
+
### Dataset Creation
|
78 |
+
The dataset was developed using a pseudo-labeling approach.
|
79 |
+
An extensive, large-scale, and high-quality speech dataset of approximately 20,000 hours was developed for domain-agnostic Bangla ASR.
|
80 |
+
|
81 |
+
## Social Impact of Dataset
|
82 |
+
|
83 |
+
## Limitations
|
84 |
+
|
85 |
+
## Citation Information
|
86 |
+
You can access the MegaBNSpeech paper at _________________ Please cite the paper when referencing the MegaBNSpeech corpus as:
|
87 |
+
```
|
88 |
+
@article{_______________,
|
89 |
+
title = {_______________________________},
|
90 |
+
author = {___,___,___,___,___,___,___,___},
|
91 |
+
journal={_______________________________},
|
92 |
+
url = {_________________________________},
|
93 |
+
year = {2023},
|