AGROASSAM: A Web Based Assamese Speech Recognition Application for Retrieving Agricultural Commodity Price and Weather Information

Abhishek Dey, Abhash Deka, Siddika Imani, Barsha Deka, Rohit Sinha, S R Mahadeva Prasanna, Priyankoo Sarmah, K Samudravijaya, S. R. NirmalaInterspeech 2018

TL;DR

This paper introduces AGROASSAM, a web-based application that uses Assamese speech recognition to retrieve agricultural commodity prices and weather information. It demonstrates the seamless integration of existing phone-based speech recognition systems into a web platform without significant new development costs, making vital information more accessible to farmers.

What Problem Does This Paper Address?

In agrarian economies like India, timely access to agricultural commodity prices and weather forecasts is crucial for farmers. While previous efforts focused on toll-free phone systems, the increasing adoption of smartphones, tablets, and computers necessitates more accessible, web-based solutions. The challenge lies in adapting existing voice-query systems, originally designed for phone lines, to a web interface without incurring substantial costs for new system development, while also handling potential issues like channel mismatch and background noise effectively.

Methods

The core of AGROASSAM relies on an Automatic Speech Recognition (ASR) system built using the Kaldi toolkit. The system was trained on a custom Assamese speech corpus collected from various districts to capture dialectal variations. Key techniques include:

  • Feature Extraction: Mel-frequency cepstral coefficients (MFCCs) were computed from speech signals, enhanced by splicing to capture dynamic characteristics. These features were then reduced and decorrelated using Linear Discriminant Analysis (LDA) and Maximum Linear Likelihood Linear Transformations (MLLT).
  • Acoustic Modeling: Two approaches were explored:
  • GMM-HMM: A conventional context-dependent Gaussian Mixture Model Hidden Markov Model system.
  • DNN-HMM: A Deep Neural Network (DNN) based system, leveraging Karel's implementation.
  • Noise Handling: A Zero Frequency Filtered Signal (ZFFS) based technique was employed to separate foreground speech from background noise, a common issue in real-world field recordings.
  • Language Modeling: Task-specific modified language models were used to enhance recognition accuracy.

Datasets

The paper describes a custom speech corpus collected for this work, comprising:

  • Isolated Commodity and District Names: 30 hours of data from 885 speakers, covering 109 commodity names and 27 district names.
  • Phonetically Balanced Sentences: 3 hours of data from 25 speakers.
  • Continuous Speech: An additional 5658 speech files from 27 speakers.

For evaluation, two test sets were used:

  • Commodity Test Set: 2552 utterances of isolated commodity names.
  • District Test Set: 1125 utterances of isolated district names.

Key Insights and Ideas

  • Seamless Web Integration: The most significant insight is that existing voice-query systems designed for phone lines can be effectively and seamlessly integrated into a web-based platform with minimal additional development effort. This opens up possibilities for repurposing existing ASR infrastructure.
  • Dual-Use Systems: The same system can simultaneously serve both phone-line and web-interface users without performance degradation, maximizing resource utilization.
  • Accessibility for Farmers: By leveraging speech recognition, AGROASSAM makes critical agricultural information accessible to farmers in their native Assamese language, bridging the digital divide and supporting an agrarian economy.
  • Handling Real-World Conditions: The successful application of noise reduction techniques (ZFFS) and the use of diverse speech data demonstrate the system's robustness in handling the challenges of real-world field conditions.

Key Results

The paper reports the following Word Error Rates (WERs) for the two acoustic modeling approaches:

  • GMM-HMM:
    • Commodity: 10.31%
    • District: 5.16%
  • DNN-HMM:
    • Commodity: 7.79%
    • District: 4.98%

The DNN-HMM system achieved relative improvements of 24.44% in commodity name recognition and 3.48% in district name recognition compared to the GMM-HMM system.

Limitations and Future Work

  • Limitations: The paper doesn't explicitly detail limitations beyond the reported WERs. However, the reliance on specific data collection methods and the focus on Assamese language suggest potential limitations in generalizability to other languages or broader domains without further adaptation.
  • Future Work: The authors suggest that the demonstrated web-based systems can be implemented on public kiosks to further enhance information accessibility for users in their native languages. The success of integrating existing phone-based systems into a web platform also implies potential for further optimization and expansion to other languages and information domains.

Summary updated

BibTeX
@inproceedings{dey2018agroassam,
  author = {Abhishek Dey and Abhash Deka and Siddika Imani and Barsha Deka and Rohit Sinha and S R Mahadeva Prasanna and Priyankoo Sarmah and K Samudravijaya and S. R. Nirmala},
  title = {{AGROASSAM: A Web Based Assamese Speech Recognition Application for Retrieving Agricultural Commodity Price and Weather Information}},
  booktitle = {Interspeech},
  year = {2018},
}
Figure