Voice search for development

Etienne Barnard, Johan Schalkwyk, Charl van Heerden, Pedro J. MorenoInterspeech 2010

TL;DR

This paper argues that voice search, a powerful tool for information access in developed nations, holds immense potential to bridge the digital divide in developing countries. By overcoming challenges related to illiteracy, cost, and content availability, voice search can democratize access to information and empower developing-world citizens.

What Problem Does This Paper Address?

Despite the widespread belief that speech technology can significantly improve the lives of people in developing countries, its impact has been limited. This is largely due to challenges such as widespread illiteracy, high costs of information access (e.g., expensive data plans and telephone calls), and a scarcity of relevant digital content. Existing speech technology applications often mirror those in developed countries, using directed dialogs or kiosk-like interfaces, which don't fully address the unique needs and constraints of developing regions. The authors propose voice search as a novel solution that can overcome these barriers.

Methods

The paper proposes voice search as a solution by leveraging its inherent characteristics: free-form input, massive language models, and the ability to combine spoken and visual information. The authors outline the technical challenges in developing voice search for under-resourced languages, focusing on acoustic and language modeling. They discuss strategies for data collection and model training, acknowledging the need for adaptations due to linguistic diversity (e.g., tone languages, click sounds). For language modeling, they explore methods for languages with limited textual resources, including mining existing text, using translations, and leveraging proper names. They also consider the practicalities of user interfaces, suggesting speech-only interactions and the use of text-to-speech for users with limited handset capabilities or literacy.

Datasets

The paper describes the use of text queries from anonymized computer users who selected Zulu or Afrikaans as their browser language. These queries were used to create prompting materials for speech data collection. Approximately 200,000 spoken queries were collected for each language. The authors also mention using a manually crafted finite-state grammar to generate additional Zulu prompts. For evaluation, they used held-out test sets of these text queries and spoken utterances.

Key Insights and Ideas

  • Voice Search as an Equalizer: Voice search can bypass literacy barriers and high data costs, making information accessible to a much wider population in developing countries.
  • Overcoming Content Scarcity: Voice search can access web content regardless of its original source, mitigating the problem of fragmented and scarce digital information.
  • Flexible Technology Appropriation: Similar to how "missed calls" have been innovatively used in developing countries, voice search's flexibility allows local communities to adapt its usage to their specific social contexts.
  • Multilingualism is Key: Many developing countries are highly multilingual. Voice search systems need to handle code-switching between local languages and dominant "world" or regional languages, and acoustic/language models should be tailored to accented speech.
  • Proper Names are Crucial: The pronunciation of proper names varies significantly, posing a challenge for accurate recognition. Understanding these variations is vital for robust voice search.
  • Beyond Visual Displays: To reach users with limited handset capabilities or literacy, voice-only interfaces with text-to-speech output are essential.

Key Results

The paper presents initial voice search results for Afrikaans and Zulu. Their systems achieved normalized sentence accuracy (NSACC) of 67.0% for Afrikaans and 65.4% for Zulu, and web score top 1 (WSC1) accuracies of 71.9% and 69.3% respectively. These results are noted as being 20-50% higher in error rates compared to a US English baseline, but are considered encouraging given the limited training data and the potential for user acceptance. The results also suggest that a smaller amount of training data (around 90,000 utterances) might be sufficient for initial releases.

Limitations and Future Work

The authors acknowledge that their current voice search implementations typically rely on visual displays for search results, limiting usability for users without feature-rich handsets or literacy. They also note that many current systems depend heavily on data communication, which might not be universally available. Future work will involve further investigation into these practical issues, including developing speech-only interfaces and exploring designs that utilize only the voice channel for communication. They also plan to extend their development to more languages and conduct more detailed usage studies.

Summary updated

BibTeX
@inproceedings{barnard2010voice,
  author = {Etienne Barnard and Johan Schalkwyk and Charl van Heerden and Pedro J. Moreno},
  title = {{Voice search for development}},
  booktitle = {Interspeech},
  year = {2010},
}
Figure