Automatic speech recognition for under-resourced languages: A survey

A method for operating a voice trigger is provided. In some implementations, the method is performed at an electronic device including one or more processors and memory storing instructions for execution by the one or more processors. The method includes receiving a sound input. The sound input may correspond to a spoken word or phrase, or a portion thereof. The method includes determining whether at least a portion of the sound input corresponds to a predetermined type of sound, such as a human voice. The method includes, upon a determination that at least a portion of the sound input corresponds to the predetermined type, determining whether the sound input includes predetermined content, such as a predetermined trigger word or phrase. The method also includes, upon a determination that the sound input includes the predetermined content, initiating a speech-based service, such as a voice-based digital assistant.

Voice trigger for a digital assistant

We present a simple method for transferring dependency parsers from source languages with labeled training data to target languages without labeled training data. We first demonstrate that delexicalized parsers can be directly transferred between languages, producing significantly higher accuracies than unsupervised parsers. We then use a constraint driven learning algorithm where constraints are drawn from parallel corpora to project the final parser. Unlike previous work on projecting syntactic resources, we show that simple methods for introducing multiple source languages can significantly improve the overall quality of the resulting parsers. The projected parsers from our system result in state-of-the-art performance when compared to previously studied unsupervised and projected parsing systems across eight different languages.

/pdf/multi-source-transfer-of-delexicalized-dependency-parsers-zp99s8yy8y.pdf

Multi-Source Transfer of Delexicalized Dependency Parsers

https://danielpovey.com/files/csl10_sgmm_preprint.pdf

The subspace Gaussian mixture model-A structured model for speech recognition

In this work, a novel training scheme for generating bottleneck features from deep neural networks is proposed. A stack of denoising auto-encoders is first trained in a layer-wise, unsupervised manner. Afterwards, the bottleneck layer and an additional layer are added and the whole network is fine-tuned to predict target phoneme states. We perform experiments on a Cantonese conversational telephone speech corpus and find that increasing the number of auto-encoders in the network produces more useful features, but requires pre-training, especially when little training data is available. Using more unlabeled data for pre-training only yields additional gains. Evaluations on larger datasets and on different system setups demonstrate the general applicability of our approach. In terms of word error rate, relative improvements of 9.2% (Cantonese, ML training), 9.3% (Tagalog, BMMI-SAT training), 12% (Tagalog, confusion network combinations with MFCCs), and 8.7% (Switchboard) are achieved.

/pdf/extracting-deep-bottleneck-features-using-stacked-auto-5dqrqmtrxv.pdf

Extracting deep bottleneck features using stacked auto-encoders

Systems and methods for spoken term detection are provided. A method for spoken term detection, comprises receiving phone level out-of-vocabulary (OOV) keyword queries, converting the phone level OOV keyword queries to words, generating a confusion network (CN) based keyword searching (KWS) index, and using the CN based KWS index for both in-vocabulary (IV) keyword queries and the OOV keyword queries.

Method and system for efficient spoken term detection using confusion networks

In this paper, we describe the IBM MASTOR, a speech-to-speech translation system that can translate spontaneous free-form speech in real-time on both laptop and hand-held PDAs. Challenges include speech recognition and machine translation in adverse environments, lack of training data and linguistic resources for under-studied languages, and the need to rapidly develop capabilities for new languages. Another challenge is designing algorithms and building models in a scalable manner to perform well even on memory and CPU deficient hand-held computers. We describe our approaches, experience, and success in building working free-form S2S systems that can handle two language pairs (including a low-resource language).

IBM MASTOR SYSTEM: Multilingual Automatic Speech-to-Speech Translator

The paper describes a state-of-the-art spoken term detection system in which significant improvements are obtained by diversifying the ASR engines used for indexing and combining the search results. First, we describe the design factors that, when varied, produce complementary STD systems and show that the performance of the combined system is 3 times better than the best individual component. Next, we describe different strategies for system combination and show that significant improvements can be achieved by normalizing the combined scores. We propose a classifier-based system combination strategy which outperforms a highly optimized baseline. The system described in this paper had the highest accuracy in the 2012 DARPA RATS evaluation.

Exploiting diversity for spoken term detection

This paper describes the Arabic broadcast transcription system fielded by IBM in the GALE Phase 2.5 machine translation evaluation. Key advances include the use of additional training data from the Linguistic Data Consortium (LDC), use of a very large vocabulary comprising 737 K words and 2.5 M pronunciation variants, automatic vowelization using flat-start training, cross-adaptation between unvowelized and vowelized acoustic models, and rescoring with a neural-network language model. The resulting system achieves word error rates below 10% on Arabic broadcasts. Very large scale experiments with unsupervised training demonstrate that the utility of unsupervised data depends on the amount of supervised data available. While unsupervised training improves system performance when a limited amount (135 h) of supervised data is available, these gains disappear when a greater amount (848 h) of supervised data is used, even with a very large (7069 h) corpus of unsupervised data. We also describe a method for modeling Arabic dialects that avoids the problem of data sparseness entailed by dialect-specific acoustic models via the use of non-phonetic, dialect questions in the decision trees. We show how this method can be used with a statically compiled decoding graph by partitioning the decision trees into a static component and a dynamic component, with the dynamic component being replaced by a mapping that is evaluated at run-time.

Advances in Arabic Speech Transcription at IBM Under the DARPA GALE Program

In a speech recognition system, the combination of a log-linear model with a multitude of speech features is provided to recognize unknown speech utterances. The speech recognition system models the posterior probability of linguistic units relevant to speech recognition using a log-linear model. The posterior model captures the probability of the linguistic unit given the observed speech features and the parameters of the posterior model. The posterior model may be determined using the probability of the word sequence hypotheses given a multitude of speech features. Log-linear models are used with features derived from sparse or incomplete data. The speech features that are utilized may include asynchronous, overlapping, and statistically non-independent speech features. Not all features used in training need to appear in testing/recognition.

Hong-Kwang Kuo

Papers

Method and system for efficient spoken term detection using confusion networks

IBM MASTOR SYSTEM: Multilingual Automatic Speech-to-Speech Translator

Exploiting diversity for spoken term detection

Advances in Arabic Speech Transcription at IBM Under the DARPA GALE Program

Speech recognition utilizing multitude of speech features