Hugging Face's logo

Datasets:
SEACrowd
/
indspeech_teldialog_svcsr

Languages:
Indonesian
Tags:
speech-recognition
Dataset card Files Community
Dataset Viewer
Viewer

The viewer is disabled because this dataset repo requires arbitrary Python code execution. Please consider removing the loading script and relying on automated data support (you can use convert_to_parquet from the datasets library). If this is not possible, please open a discussion for direct help.

indspeech_teldialog_svcsr

This is the first Indonesian speech dataset for small vocabulary continuous speech recognition (SVCSR).

The data was developed by TELKOMRisTI (R&D Division, PT Telekomunikasi Indonesia) in collaboration with Advanced

Telecommunication Research Institute International (ATR) Japan and Bandung Institute of Technology (ITB) under the

Asia-Pacific Telecommunity (APT) project in 2004 [Sakti et al., 2004]. Although it was originally developed for

a telecommunication system for hearing and speaking impaired people, it can be used for other applications,

i.e., automatic call centers. Furthermore, as all speakers utter the same sentences,

it can also be used for voice conversion tasks.

The text is based on a word vocabulary which is derived from some necessary dialog calls,

such as dialog calls with the 119 emergency department, 108 telephone information department,

and ticket reservation department. In total, it consists of 20,000 utterances (about 18 hours of speech) from the

70-word dialog vocabulary of 100 sentences (including single word sentences) each uttered by 200 speakers

(100 Females, 100 Males). The age is limited to middle age (20-40 years), but they present a wide range of spoken

dialects from different ethnic groups. The recording is conducted in parallel for both clean and telephone speech,

but we open only the clean speech due to quality issues on telephone speech.

Each audio file is a single-channel 16-bit PCM WAV with a sample rate of 16000 Hz.

These utterances are equally split into training and test sets with 100 speakers (50 Females, 50 Males) in each set.

Dataset Usage

Run pip install nusacrowd before loading the dataset through HuggingFace's load_dataset.

Citation 

@inproceedings{sakti-icslp-2004,
    title = "Indonesian Speech Recognition for Hearing and Speaking Impaired People",
    author = "Sakti, Sakriani and Hutagaol, Paulus and Arman, Arry Akhmad and Nakamura, Satoshi",
    booktitle = "Proc. International Conference on Spoken Language Processing (INTERSPEECH - ICSLP)",
    year = "2004",
    pages = "1037--1040"
    address = "Jeju Island, Korea"
}

License

CC-BY-NC-SA-4.0

Homepage

https://github.com/s-sakti/data_indsp_teldialog_svcsr/

NusaCatalogue

For easy indexing and metadata: https://indonlp.github.io/nusa-catalogue

Downloads last month
0
Edit dataset card