TL;DR
This paper introduces a novel discriminative training (DT) method for keyword spotting (KWS) on spontaneous speech. By adapting Minimum Classification Error (MCE) to assign non-uniform error costs, the approach significantly improves keyword detection accuracy, especially in challenging conversational scenarios.
What Problem Does This Paper Address?
Recognizing speech in spontaneous conversations is notoriously difficult due to relaxed speaking styles and less structured language compared to dictated speech. Traditional Automatic Speech Recognition (ASR) systems, optimized for high word accuracy, often struggle in these scenarios. Keyword spotting (KWS) offers a solution by focusing on identifying specific keywords rather than transcribing entire utterances, which is crucial for tasks like call routing. However, existing KWS methods often lack a unified framework for optimizing keyword detection accuracy, especially when dealing with the inherent complexities of spontaneous speech and the need to balance false alarms with missed keywords. This paper aims to bridge this gap by proposing a discriminative training approach that specifically prioritizes keyword accuracy.
Methods
The core contribution is the adaptation of Minimum Classification Error (MCE), a discriminative training technique, for keyword spotting. MCE typically aims to minimize the overall word error rate. Here, the authors extend it to Non-uniform MCE, where different types of errors are assigned different costs. Specifically, they introduce a non-uniform error cost function that assigns higher costs to errors related to keywords.
This non-uniform MCE is implemented within the Weighted Finite State Transducer (WFST) framework. This is a powerful approach for speech recognition that allows for the efficient representation and manipulation of acoustic models, lexicons, and grammars. The authors leverage WFST's difference operation to effectively exclude the reference transcription from the hypothesis space during training, a crucial step for MCE. The error cost function is designed to be time-dependent, allowing for emphasis on frames identified as keywords or likely to be recognized as keywords.
Datasets
The experiments were conducted on the Switchboard-1 Release 2 corpus, a collection of spontaneous telephone conversations. For evaluation, a subset of conversations related to "CREDIT CARD USE" was extracted. This subset was chosen because the topic information is not readily available in the standard release, making it a more realistic testbed.
Key Insights and Ideas
- Keyword-centric Optimization: The fundamental idea is that for keyword spotting, not all errors are equal. Errors involving keywords should be penalized more heavily than others.
- WFST as a Unified Framework: The paper demonstrates the power of WFSTs not only for decoding but also for implementing advanced training techniques like non-uniform MCE, enabling efficient manipulation of lattices.
- Time-Dependent Error Costs: The ability to assign error costs based on the time frame within an utterance is key to focusing training on relevant speech segments.
- Two Scenarios, One Solution: The proposed method shows effectiveness in both scenarios: with a full Large Vocabulary Continuous Speech Recognition (LVCSR) decoder and with a simpler word-loop grammar, highlighting its robustness.
Key Results
The non-uniform MCE approach consistently outperformed the baseline and standard MCE methods:
- LVCSR Decoder Scenario: Achieved a 2.77% absolute improvement in Figure of Merit (FOM) over the baseline and a 1.02% improvement over standard MCE.
- Word-Loop Grammar Scenario: Achieved a 3.15% absolute improvement in FOM over the baseline and a 2.16% improvement over standard MCE.
The ROC curves also indicated that non-uniform MCE achieved higher detection rates across various False Alarm (FA) rates in both scenarios.
Limitations and Future Work
- Error Cost Function Design: The authors acknowledge that the current error cost function is relatively simple. They suggest that more sophisticated designs, potentially incorporating context-dependent costs or asymmetrical penalties for misses versus false alarms, could lead to further performance gains.
- Computational Overhead: While the WFST implementation is claimed to be efficient, the paper doesn't deeply explore the computational cost compared to standard MCE or MLE training, especially for very large datasets or complex error cost functions.
- Future Directions: The authors explicitly mention exploring more sophisticated error cost functions as a primary avenue for future work. They also hint at the potential for applying this approach to other related tasks within speech understanding.
Summary updated