Abstract
Tracter is a lightweight, pull-driven C++ framework designed to simplify the construction of complex signal processing pipelines for Automatic Speech Recognition (ASR). By treating processing steps as nodes in a directed graph, it allows researchers to seamlessly integrate disparate third-party libraries into unified, real-time-capable systems.
Introduction
In the domain of speech and image processing, researchers frequently encounter "pipeline rot." A typical ASR system involves a long chain of operations: sampling, filtering, spectral analysis, and feature extraction. While implementing a single step is straightforward, embedding that step into a large, real-time system—especially one involving collaborative work across different institutions—is notoriously difficult.
Existing tools often force a choice between rigid, high-performance "push" systems (where data is shoved through the pipe) and flexible but slow offline scripts. Tracter addresses this need by providing a framework that handles both online (live) and offline (file-based) processing while offering a standardized way to wrap specialized libraries (like FFT or machine learning toolkits) into a cohesive graph.
The Dataflow Paradigm
Tracter adopts the dataflow paradigm, where individual components (actors) are linked as vertices in a directed graph. This structure allows for modular development and easy reconfiguration of processing steps.
Pull-Driven Marshalling
Unlike many real-time systems that "push" data from the source, Tracter is "pull-driven." The "sink" (usually the speech decoder) requests data, and this request propagates upstream through the graph. This decoder-centric approach ensures that processing is only performed on data that is actually required, optimizing computational efficiency.
Automated Buffer Management
A key technical feature of Tracter is its message-passing algorithm for cache sizing. At construction time, the sink informs its inputs of the amount of data needed, including "look-ahead" or "look-behind" offsets. This information ripples through the graph, allowing every component to allocate a fixed-size circular buffer exactly once. This avoids costly runtime memory allocation and simplifies the implementation of features requiring temporal context, such as Delta and Acceleration features.
Integration and Flexibility
Tracter acts as a "wrapper framework," providing interfaces for popular libraries like FFTW (for Fourier transforms), Torch (for neural networks), BSAPI, and HTK. This allows researchers to leverage specialized tools without reinventing core functionalities.
One elegant application of this flexibility is the "gate" pattern for Voice Activity Detection (VAD). The authors demonstrate how VAD can be handled as a gate component in the graph that only allows features to pass to the decoder when speech is detected, effectively pausing the expensive decoding process during silence.
Key Results
The framework's utility is demonstrated through its ability to integrate heterogeneous tools into complex processing graphs. Tracter was developed within the AMI/AMIDA consortium and serves as the primary feature extraction engine for the Juicer ASR decoder.
The framework successfully bridges the gap between offline research and real-time deployment, allowing the exact same code to be used in both environments. Furthermore, Tracter has been released under a permissive BSD-style license, making it a viable tool for both academic and commercial applications.
Limitations and Future Work
Despite its strengths, Tracter currently lacks internal parallelism; components run serially within a single graph. Additionally, the VAD design introduces a small amount of latency due to necessary look-ahead, which may be a hurdle for ultra-low-latency interactive systems. Future work may focus on optimizing buffer allocation in complex branching graphs and exploring multi-core exploitation within the pipeline.
Summary updated