Abstract
We study the segmental recurrent neural network for end-to-end acoustic modelling. This model connects the segmental conditional random field (CRF) with a recurrent neural network (RNN) used for feature extraction. Compared to most previous CRF-based acoustic models, it does not rely on an external system to provide features or segmentation boundaries. Instead, this model marginalises out all the possible segmentations, and features are extracted from the RNN trained together with the segmental CRF. In essence, this model is self-contained and can be trained end-to-end. In this paper, we discuss practical training and decoding issues as well as the method to speed up the training in the context of speech recognition. We performed experiments on the TIMIT dataset. We achieved 17.3 phone error rate (PER) from the first-pass decoding --- the best reported result using CRFs, despite the fact that we only used a zeroth-order CRF and without using any language model.
Introduction
Traditional speech recognition systems have long relied on Hidden Markov Models (HMMs). However, HMMs are limited by two significant drawbacks: the "independence assumption," which posits that acoustic frames are independent of one another, and the requirement for complex, multi-stage pipelines that are difficult to optimize jointly.
While modern end-to-end alternatives like Connectionist Temporal Classification (CTC) and Attention-based models have emerged, they introduce new challenges. CTC still performs frame-level classification and requires input/output lengths to match. Attention-based models, though powerful, do not naturally account for the linear and local alignment of speech signals, where a phoneme corresponds to a specific temporal window rather than a global summary of the utterance. This paper introduces the Segmental Recurrent Neural Network (SRNN) to address these limitations by handling variable-length segments naturally and learning feature representations end-to-end.
The Segmental RNN Architecture
The core of the SRNN is the integration of a segmental Conditional Random Field (CRF) with an RNN encoder. Unlike standard CRFs that label every frame, a segmental CRF defines probabilities over segments of time, allowing the model to hypothesize boundaries for specific phonemes.
As shown in the architecture above, a bi-directional LSTM processes raw acoustic features. For any proposed segment defined by a start time $s$ and end time $e$, a separate RNN extracts a fixed-dimensional "segment embedding." Because the true boundaries are often unknown, the model employs a dynamic programming algorithm to sum over all possible segmentations, learning both boundaries and features simultaneously.
Computational Efficiency via Subsampling
A primary challenge of segmental models is the $O(T^2)$ computational complexity required to search all possible segments. To mitigate this, the authors implement hierarchical subsampling layers within the RNN. By skipping frames or concatenating hidden states, the model effectively shortens the sequence length.
As demonstrated in the table above, adding two layers of subsampling reduces the maximum segment length search space and achieves a 10x speedup in training and decoding with negligible impact on accuracy. Interestingly, the authors found that "skipping" frames in lower layers actually improved performance by reducing the inherent redundancy in high-frame-rate audio signals.
Experimental Results
The SRNN was evaluated on the TIMIT dataset, a standard benchmark for phone recognition. The model achieved a 17.3% Phone Error Rate (PER) on the core test set. This result is significant as it:
- Outperforms previous segmental CRF approaches (typically 19-21%).
- Is highly competitive with other end-to-end methods like CTC (18.4%) and Attention-based RNNs (17.6%).
- Achieves state-of-the-art performance for CRF-based models without the aid of an external language model.
Conclusion and Future Work
The SRNN represents a "self-contained" approach to speech recognition, discovering phoneme boundaries through marginalization rather than relying on pre-segmented data from an external HMM. While the $O(T^2)$ nature of the model remains more demanding than frame-level models, hierarchical subsampling makes it a viable candidate for end-to-end tasks. Future research will focus on scaling the model to large-vocabulary tasks, implementing beam search to reduce decoding errors, and integrating language models via weighted finite-state transducers (WFST).
Summary updated