Abstract
This paper details the "Patrol Team" submission for the first phase of the DARPA Robust Automatic Transcription of Speech (RATS) program. The team, a collaboration between Brno University of Technology, Raytheon BBN, and MIT CSAIL, developed a Language Identification (LID) system designed for extremely noisy and degraded radio communication environments. By fusing acoustic and phonotactic i-vector systems and employing neural network classifiers, the team achieved significant performance gains across various audio durations, particularly in the challenging 3-second condition.
Introduction
Traditional Language Identification (LID) research has primarily focused on relatively clean telephone or broadcast data. However, real-world signals in defense and emergency contexts are often transmitted over low-quality radio channels, characterized by non-additive noise and "push-to-talk" (PTT) clipping. The DARPA RATS program was established to address these challenges, requiring the identification of five target languages—Dari, Arabic Levantine, Urdu, Pashto, and Farsi—from audio segments as short as 3 seconds recorded over eight highly degraded communication channels.
System Architecture
The Patrol Team utilized a fusion-based approach, integrating four distinct subsystems to maximize robustness:
- Acoustic JFA: A Joint Factor Analysis model using shifted-delta-cepstra (SDC) features, pre-processed with Wiener filtering to suppress channel noise.
- i-vector NN (BUT): An acoustic i-vector extractor where classification is performed by a three-layer Neural Network (NN), which proved more effective than standard logistic regression for noisy data.
- i-vector NN (BBN): This system utilized Language-Dependent Background Models (LDBM) and was trained on 20-second audio "chunks" to better simulate the short-duration segments encountered during evaluation.
- Phonotactic (PHN-CZ): A system that converts speech into phoneme "soft-counts" using a Czech phone recognizer, which are then modeled using i-vectors.
The final system output is a weighted fusion of these four subsystems, calibrated to produce log-likelihood ratios.
Key Technical Insights
Several strategies were critical to the system's success in the RATS environment. Applying a Wiener filter as a pre-processor provided a consistent boost in performance, particularly for acoustic models. Furthermore, the use of Neural Networks for classification and the adoption of Language-Dependent Background Models (LDBM) were pivotal.
As shown in the table above, the transition from a standard Universal Background Model (UBM) to an LDBM, combined with appropriate chunking (20s), led to substantial reductions in $C_{avg}$, particularly for shorter audio segments.
Evaluation and Results
The system was evaluated using the average cost of detection ($C_{avg}$) across four test durations: 120s, 30s, 10s, and 3s.
The results demonstrate the power of system fusion. On the 120s development set, the fusion system achieved a $C_{avg}$ of 0.83%, nearly halving the error of the best individual system (i-vector BUT at 1.60%). While performance naturally degrades as audio duration decreases, the fusion system maintained a competitive $C_{avg}$ of 17.45% even in the extreme 3-second evaluation condition.
Conclusion
The Patrol Team's work confirms that state-of-the-art LID techniques like i-vectors remain highly effective in extreme noise conditions when properly adapted. The combination of Wiener filtering, neural network classification, and the fusion of acoustic and phonotactic models provides a robust framework for identifying languages in degraded communication channels. Future work will focus on further improving performance for ultra-short segments and better handling of non-speech regions in radio transmissions.
Summary updated