TIMIT

About: TIMIT is a research topic. Over the lifetime, 1401 publications have been published within this topic receiving 59888 citations. The topic is also known as: TIMIT Acoustic-Phonetic Continuous Speech Corpus.

When CTC Training Meets Acoustic Landmarks

Abstract: Connectionist temporal classification (CTC) provides an end-to-end acoustic model (AM) training strategy. CTC learns accurate AMs without time-aligned phonetic transcription, but sometimes fails to converge, especially in resource-constrained scenarios. In this paper, the convergence properties of CTC are improved by incorporating acoustic landmarks. We tailored a new set of acoustic landmarks to help CTC training converge more rapidly and smoothly while also reducing recognition error rates. We leveraged new target label sequences mixed with both phone and manner changes to guide CTC training. Experiments on TIMIT demonstrated that CTC based acoustic models converge significantly faster and smoother when they are augmented by acoustic landmarks. The models pretrained with mixed target labels can be further finetuned, resulting in phone error rates 8.72% below baseline on TIMIT. Consistent performance gain is also observed on WSJ (a larger corpus) and reduced TIMIT (smaller). With WSJ, we are the first to succeed in verifying the effectiveness of acoustic landmark theory on a mid-sized ASR task.

Evaluation of sentence list equivalency for the TIMIT sentences by cochlear implant recipients

Abstract: The equivalency of 34 TIMIT sentence lists was evaluated using adult cochlear implant recipients to determine if they should be recommended for future clinical or research use. Because these sentences incorporate gender, dialect and speaking rate variations, they have the potential to better represent speech recognition abilities in real-world communication situations. Copyright by: Sarah E. King May 2008

High Level Speaker Specific Features as an Efficiency Enhancing Parameters in Speaker Recognition System

Abstract: In this paper, I present high-level speaker specific feature extraction considering intonation, linguistics rhythm, linguistics stress, prosodic features directly from speech signals. I assume that the rhythm is related to language units such as syllables and appears as changes in measurable parameters such as fundamental frequency ( ), duration, and energy. In this work, the syllable type features are selected as the basic unit for expressing the prosodic features. The approximate segmentation of continuous speech to syllable units is achieved by automatically locating the vowel starting point. The knowledge of high-level speaker’s specific speakers is used as a reference for extracting the prosodic features of the speech signal. High-level speaker-specific features extracted using this method may be useful in applications such as speaker recognition where explicit phoneme/syllable boundaries are not readily available. The efficiency of the particular characteristics of the specific features used for automatic speaker recognition was evaluated on TIMIT and HTIMIT corpora initially sampled in the TIMIT at 16 kHz to 8 kHz. In summary, the experiment, the basic discriminating system, and the HMM system are formed on TIMIT corpus with a set of 48 phonemes. Proposed ASR system shows 1.99%, 2.10%, 2.16% and 2.19 % of efficiency improvements compared to traditional ASR system for and of 16KHz TIMIT utterances.

Online learning of large margin hidden markov models for automatic speech recognition

Abstract: Over the last two decades, large margin methods have yielded excellent performance on many tasks. The theoretical properties of large margin methods have been intensively studied and are especially well-established for support vector machines (SVMs). However, the scalability of large margin methods remains an issue due to the amount of computation they require. This is especially true for applications involving sequential data. In this thesis we are motivated by the problem of automatic speech recognition (ASR) whose large-scale applications involve training and testing on extremely large data sets. The acoustic models used in ASR are based on continuous-density hidden Markov models (CD-HMMs). Researchers in ASR have focused on discriminative training of HMMs, which leads to models with significantly lower error rates. More recently, building on the successes of SVMs and various extensions thereof in the machine learning community, a number of researchers in ASR have also explored large margin methods for discriminative training of HMMs. This dissertation aims to apply various large margin methods developed in the machine learning community to the challenging large-scale problems that arise in ASR. Specifically, we explore the use of sequential, mistake-driven updates for online learning and acoustic feature adaptation in large margin HMMs. The updates are applied to the parameters of acoustic models after the decoding of individual training utterances. For large margin training, the updates attempt to separate the log-likelihoods of correct and incorrect transcriptions by an amount proportional to their Hamming distance. For acoustic feature adaptation, the updates attempt to improve recognition by linearly transforming the features computed by the front end. We evaluate acoustic models trained in this way on the TIMIT speech database. We find that online updates for large margin training not only converge faster than analogous batch optimizations, but also yield lower phone error rates than approaches that do not attempt to enforce a large margin. We conclude this thesis with a discussion of future research directions, highlighting in particular the challenges of scaling our approach to the most difficult problems in large-vocabulary continuous speech recognition.

Single Channel multi-speaker speech Separation based on quantized ratio mask and residual network

Abstract: The recently-proposed deep clustering-based algorithms represent a fundamental advance towards the single-channel multi-speaker speech sep- aration problem. These methods use an ideal binary mask to construct the objective function and K-means clustering method to estimate the ideal bina- ry mask. However, when sources belong to the same class or the number of sources is large, the assumption that one time-frequency unit of the mixture is dominated by only one source becomes weak, and the IBM-based separation causes spectral holes or aliasing. Instead, in our work, the quantized ideal ratio mask was proposed, the ideal ratio mask is quantized to have the output of the neural network with a limited number of possible values. Then the quan- tized ideal ratio mask is used to construct the objective function for the case of multi-source domination, to improve network performance. Furthermore, a network framework that combines a residual network, a recurring network, and a fully connected network was used for exploiting correlation information of frequency in our work. We evaluated our system on TIMIT dataset and show 1.6 dB SDR improvement over the previous state-of-the-art methods.