Source-linked AI summary
Google Crowdsourced Speech Corpora and Related Open-Source Resources for Low-Resource Languages and Dialects: An Overview
Alena Butryna, Shan-Hui Cathy Chu, Isin Demirsahin, Alexander Gutkin, Linne Ha, Fei He, Martin Jansche, Cibu Johny, Anna Katanova, Oddur Kjartansson, Chenfang Li, Tatiana Merkulova, Yin May Oo, Knot Pipatsrisawat, Clara Rivera, Supheakmungkol Sarin, Pasindu de Silva, Keshan Sodimana, Richard Sproat, Theeraphol Wattanavekin, Jaka Aris Eko Wibawa
TL;DR
Under-represented languages lack freely available speech resources, despite the need for ASR and TTS resources and methods suited to local contexts. The paper surveys a program that collects and releases corpora through community partnerships and varied collection practices. It reports 38 released datasets and more than 1,500 hours of speech, while identifying linguistic-resource development as a continuing scaling challenge.
Problem
Many endangered and low-resource languages lack the linguistic resources needed for speech technology, and only an estimated 5% of the world’s languages are accessible on the Internet.
Source: §1 · p. 1Method
The program selects languages across regions and language families, develops ASR and TTS corpora using public text and locally supported collection, and releases the resources freely.
Source: §2.1 · pp. 1–2, §2.2 · p. 2Results
The program released 38 datasets for ASR and TTS across four world regions, with corpora totaling over 1,500 hours of speech.
Source: §2.1 · p. 2, §3 · p. 4Takeaways & Limitations
Local university and community partnerships were crucial to program success and helped collect high-quality data.
Source: §3 · p. 4Scaling remains challenging because language-specific resource needs and scarce linguistic expertise leave no universal development recipe.
Source: §2.3 · p. 4
Abstract
from arXiv · showhide
This paper presents an overview of a program designed to address the growing need for developing freely available speech resources for under-represented languages. At present we have released 38 datasets for building text-to-speech and automatic speech recognition applications for languages and dialects of South and Southeast Asia, Africa, Europe and South America. The paper describes the methodology used for developing such corpora and presents some of our findings that could benefit under-represented language communities.
1. Introduction
The paper addresses the shortage of freely available speech resources for under-represented languages, including endangered languages and low-resource dialects. It outlines a program to develop reusable resources and methods for ASR and TTS.
Motivation
- Roughly 40% of the world’s 5,000 to 7,000 languages are classified as endangered, and an estimated 5% are accessible on the Internet.Many additional languages have large speaker populations but are low-resource because linguistic resources are lacking.Source: §1 · p. 1
Program goals
- The program aims to provide free speech resources in regions with growing Internet communities and few publicly available linguistic resources.It also seeks a replicable, open-source methodology local technologists can use to build applications or resources for new languages.Source: §1 · p. 1
Resource needs
- ASR typically needs more data from many speakers, while TTS needs higher-quality recordings from fewer speakers with well-articulated speech.Dominant-dialect technology may not serve communities whose dialects diverge enough to be mutually unintelligible.Source: §1 · p. 1
2. Program Overview
The program selects languages and dialects to broaden open-source resource availability and test collection methods across diverse linguistic settings. Its approach combines local collaboration, tailored recording and text practices, and lessons about model development and scaling challenges.
2.1. Selection of Languages and Dialects
- Language selection targets important low-resource regions and diverse language families; released TTS and ASR corpora exceed 1,500 hours and are freely available under an unencumbered license.The program includes regional languages and dialects across South and Southeast Asia, Africa, Europe, and South America.Source: §2.1 · pp. 1–2
2.2. Methodology
- Local universities and communities helped recruit volunteers, manage collection, and provide recording environments across several regions.The same collaborative blueprint was used in multiple countries, while local Google Developer group representatives supported work in Latin America.Source: §2.2 · p. 2
- ASR scripts use randomly selected Wikipedia sentences when available, while TTS scripts add handcrafted phonetic coverage, template sentences, and application-domain material.TTS material includes named entities and numeric expressions as well as examples from navigation, sports, and weather.Source: §2.2 · pp. 2–3
- ASR recordings use consumer smartphones and can be stored locally until an Internet connection is available; TTS sessions use guidance and trained native-speaker quality control.Quality checks assess transcript match, consistent volume, noise, fluency, and unnatural pauses.Source: §2.2 · p. 3
2.3. Emerging Lessons
- Moderately priced recording equipment produced corpora of sufficient quality for local communities and use in Google products.The authors also report positive effects of local collaboration, including support for locally organized crowdsourcing mechanisms and related community activities.Source: §2.3 · p. 3
- Combining crowdsourced multilingual corpora improved TTS model quality, and the South Asian model synthesized Odia and Punjabi without training data for those languages.The program also built a Malayo-Polynesian system from Javanese and Sundanese corpora combined with proprietary Indonesian data.Source: §2.3 · p. 3
- Scaling remains difficult because language-specific challenges, missing resources, and scarce linguistic expertise leave no universal development recipe.End-to-end systems and transfer learning offer potential workarounds, but can be data hungry and still benefit from linguistic knowledge.Source: §2.3 · p. 4
3. Concluding Remarks
The program released 38 datasets for TTS and ASR across languages and dialects in four world regions. The authors describe local partnerships as crucial to success, while noting that linguistic-resource development remains a challenge to scaling.
Program outcomes
- The program released 38 datasets for TTS and ASR applications across South and Southeast Asia, Africa, Europe, and South America.The collected corpora comprise over 1,500 hours of speech and are freely available online.Source: §3 · p. 4
Community role
- Partnerships with local universities and communities were crucial to program success and helped collect high-quality data.The authors hope local communities will use the methodology and datasets to build applications or collect new resources.Source: §3 · p. 4
Remaining work
- Although data collection is operationally efficient, linguistic-resource development still presents challenges to scaling the work.The authors describe the program as a foundation, with further work needed for endangered and low-resource languages.Source: §3 · p. 4