Abstract
The LeVoice system, developed by Lenovo Research for the VOiCES from a Distance Challenge 2019, addresses the significant performance degradation of Automatic Speech Recognition (ASR) in far-field environments. By integrating front-end de-reverberation, massive data augmentation, and a diverse ensemble of deep learning architectures, the system bridges the gap between clean training data and noisy real-world audio. The final system, a fusion of nine models, achieved a Word Error Rate (WER) of 19.6% on the challenge evaluation set.
Introduction
Automatic Speech Recognition (ASR) has reached high accuracy in close-talking scenarios. However, far-field speech recognition—where the speaker is distant from the microphone—remains difficult due to room reverberation and background noise. The VOiCES 2019 Challenge provided a platform to test systems against these "unmatched" conditions, using a relatively small 80-hour clean training set (Librispeech) to perform on noisy, real-world recordings. This domain mismatch is the core challenge addressed by the LeVoice system.
System Overview
The LeVoice system follows a multi-stage pipeline designed for robustness. The process begins with front-end signal processing, followed by acoustic modeling on augmented data, and concludes with decoding and system fusion.
Methods
Front-end Processing
The system utilizes the Weighted Prediction Error (WPE) algorithm for blind de-reverberation. This signal-processing technique is specifically designed to "clean" the audio before it reaches the neural network. Applying WPE provided a 10-15% relative improvement in performance, proving that traditional signal processing remains vital alongside deep learning.
Data Augmentation
To overcome the data mismatch, the 80-hour Librispeech training set was expanded 12-fold. The team applied speed perturbation and added simulated reverberation and various noises (using the RIR_NOISES dataset) to mimic the conditions of the VOiCES test set. This was the single most effective strategy for improving accuracy.
Diverse Acoustic Modeling
A variety of state-of-the-art architectures were implemented using the Kaldi toolkit to capture different acoustic features and temporal dependencies.
Key architectures included:
- CNN-TDNNF: A baseline combining Convolutional Neural Networks with Factorized Time-Delay Neural Networks.
- Residual BiLSTM: Bidirectional LSTMs with residual connections to capture long-term temporal dependencies.
- Attention Mechanisms: Self-attention layers were added to help models focus on salient audio segments.
- Phonetic Granularity: Increasing the number of phonemes used for training from 39 to 84 significantly improved the model's ability to distinguish between similar sounds in noisy environments.
System Fusion
The final results were obtained by fusing nine different models using Recognizer Output Voting Error Reduction (ROVER). This ensemble approach leverages the fact that different architectures make different types of errors, leading to better overall performance than any single model.
Results
The integration of WPE and data augmentation proved vital. For a standard TDNNF model, adding WPE reduced the WER from 24.97% to 21.67%. The best single-model configuration (CNN-TDNNF with attention and "mix-up" training) achieved a 19.82% WER on the development set. The final fused system achieved:
- Development Set: 15.91% WER
- Evaluation Set: 19.6% WER
Conclusion
The LeVoice system demonstrates that a combination of traditional signal processing (WPE), aggressive data augmentation, and diverse model ensembles can significantly mitigate the challenges of far-field ASR. Future research will focus on more advanced denoising strategies and architectures inherently robust to environmental noise.
Summary updated