addyo07/nemotron-3.5-0.6b-hinglish
Automatic Speech Recognition • Updated • 396
Datasets:
file_name stringlengths 20 41 | audio dict | transcription stringlengths 0 361 | language stringclasses 4
values | noise_type stringclasses 4
values | source stringclasses 14
values |
|---|---|---|---|---|---|
yt_train_segment_0_1.wav | {
"bytes": [
82,
73,
70,
70,
164,
163,
1,
0,
87,
65,
86,
69,
102,
109,
116,
32,
16,
0,
0,
0,
1,
0,
1,
0,
128,
62,
0,
0,
0,
125,
0,
0,
2,
0,
16,
0,
100,
97,
116,
... | To go out into that field and exchange your laundry. | en | clean | youtube_harvest |
noisy_yt_train_segment_0_1.wav | {"bytes":"UklGRqSjAQBXQVZFZm10IBAAAAABAAEAgD4AAAB9AAACABAAZGF0YYCjAQBbCnoKqwpSCw0LgAr3CdoJTQhoBzgHRA(...TRUNCATED) | To go out into that field and exchange your laundry. | en | synthesized_desktop | youtube_harvest_noisy |
yt_train_segment_0_10.wav | {"bytes":"UklGRqTDAQBXQVZFZm10IBAAAAABAAEAgD4AAAB9AAACABAAZGF0YYDDAQA6/2n/jf+J/4D/a/9A/zb/Uf+I/9n/7f(...TRUNCATED) | Independence for India was just round the corner, as Rachel mentioned. | en | clean | youtube_harvest |
noisy_yt_train_segment_0_10.wav | {"bytes":"UklGRqTDAQBXQVZFZm10IBAAAAABAAEAgD4AAAB9AAACABAAZGF0YYDDAQBRAEIAKgAbABQABgDu/9D/rv+X/4v/fP(...TRUNCATED) | Independence for India was just round the corner, as Rachel mentioned. | en | synthesized_desktop | youtube_harvest_noisy |
yt_train_segment_0_100.wav | {"bytes":"UklGRqT7AQBXQVZFZm10IBAAAAABAAEAgD4AAAB9AAACABAAZGF0YYD7AQD8//7/AAAAAAEAAAAEAAUABwALAAsACA(...TRUNCATED) | Depicts more trust, or certain brands depict more trust because. | en | clean | youtube_harvest |
yt_train_segment_0_101.wav | {"bytes":"UklGRqQLAwBXQVZFZm10IBAAAAABAAEAgD4AAAB9AAACABAAZGF0YYALAwAYABMAFwAOABMAEgAPABIAFwAbABgAHA(...TRUNCATED) | "Like this, if you do this, somebody will trust you. If you do this, somebody will trust you. Trust (...TRUNCATED) | en | clean | youtube_harvest |
noisy_yt_train_segment_0_101.wav | {"bytes":"UklGRqQLAwBXQVZFZm10IBAAAAABAAEAgD4AAAB9AAACABAAZGF0YYALAwA7ADcAMgAvAC4ALwAyADQAOQBBAEAANw(...TRUNCATED) | "Like this, if you do this, somebody will trust you. If you do this, somebody will trust you. Trust (...TRUNCATED) | en | synthesized_desktop | youtube_harvest_noisy |
yt_train_segment_0_103.wav | {"bytes":"UklGRqQvAgBXQVZFZm10IBAAAAABAAEAgD4AAAB9AAACABAAZGF0YYAvAgDY/9r/3f/Z/9b/1//Z/9n/2//Z/9z/3f(...TRUNCATED) | They have done right by the consumer, by the various stakeholders that. | en | clean | youtube_harvest |
yt_train_segment_0_104.wav | {"bytes":"UklGRqRbBABXQVZFZm10IBAAAAABAAEAgD4AAAB9AAACABAAZGF0YYBbBAACAAEAAAABAAUABwAHAA0ADQATABQAFw(...TRUNCATED) | "And therefore, today that brand commands significant trust. So I think trust is again one of those (...TRUNCATED) | en | clean | youtube_harvest |
yt_train_segment_0_106.wav | {"bytes":"UklGRqQvBABXQVZFZm10IBAAAAABAAEAgD4AAAB9AAACABAAZGF0YYAvBADp/+v/8/8AAPv/9P/w//L/7//y//3/CA(...TRUNCATED) | "As a business, you're faced with so many different decisions at some point, right? Where you can cu(...TRUNCATED) | en | clean | youtube_harvest |
End of preview. Expand in Data Studio
Noisy Hinglish ASR Corpus
A high-fidelity, mixed Hindi, English, Hinglish, and noise-robust ASR corpus designed for low-latency, localized voice assistant applications. Features conversational speech, heavy code-switching (English words embedded in Hindi structure), synthetically augmented desktop noise, and explicit non-speech negative frames for VAD optimization.
Dataset Summary
- Total Samples:
28,681audio recordings (main splits) - Total Duration:
36.6 hours - Format: Signed 16-bit PCM WAV, mono, 16,000 Hz
- Languages: Hindi (
hi), English (en), Hinglish code-switched (hinglish), Neutral non-speech (neutral) - Noise profiles:
clean,synthesized_desktop,room_background,pure_silence - Validation: 100% passed via strict structural validation
Splits
| Split | Samples | Duration | Languages |
|---|---|---|---|
| train | 23,512 | 30.26h | hi (7,726), en (6,758), hinglish (6,028), neutral (3,000) |
| val | 2,331 | 2.91h | hi (819), en (526), hinglish (536), neutral (450) |
| test | 2,838 | 3.47h | hi (832), en (504), hinglish (1,052), neutral (450) |
Benchmark Sets
| Benchmark | Samples | Duration | Description |
|---|---|---|---|
clean_hi |
500 | 37.9m | Clean high-fidelity Hindi |
hinglish |
1,036 | 62.9m | Clean code-switched Hinglish |
noisy_hi |
250 | 25.9m | Noisy/acoustic stressed Hindi |
negatives |
200 | 12.5m | Ambient noise / silence (expect empty transcription) |
Data Structure
data/
├── train/ # 23 parquet shards
├── val/ # 3 parquet shards
├── test/ # 3 parquet shards
└── benchmark/
├── clean_hi/ # 1 parquet shard
├── hinglish/ # 1 parquet shard
├── noisy_hi/ # 1 parquet shard
└── negatives/ # 1 parquet shard
Parquet Schema
| Column | Type | Description |
|---|---|---|
file_name |
string | Original WAV filename |
audio |
struct(bytes: binary, path: string) | Embedded WAV bytes + relative path |
transcription |
string | Ground truth transcript (empty for negatives) |
language |
string | hi, en, hinglish, or neutral |
noise_type |
string | clean, synthesized_desktop, room_background, pure_silence |
source |
string | Origin dataset/source |
Loading
from datasets import load_dataset
dataset = load_dataset("addyo07/noisy-hinglish-asr", streaming=True)
train = dataset["train"]
print(next(iter(train)))
Scripts
Corpus compilation and utilities are in scripts/:
compile_corpus.py— Assemble raw audio sources into structured corpuspack_to_parquet.py— Serialize WAV corpus to chunked Parquetunpack_from_parquet.py— Restore Parquet back to WAV + metadatavalidate_corpus.py— Structural validation of audio files
Related Models
This dataset was used to fine-tune addyo07/qwen-asr-0.6b-noisy-hinglish.
License
Apache 2.0
- Downloads last month
- 62