research category image

Speech recognition

Turning speech into text is at the heart of an amazing variety of products and services that enrich peoples’ lives. Most of the world’s successful speech solutions today have Nuance speech technology inside.

research category image

Speech recognition



Our goal: Near-perfect speech recognition for everybody in the world.
Nuance has been a pioneer in speech and language technologies for more than 30 years. Our data centers host billions of speech transactions every month in over 40 languages from hundreds of applications. We continuously expand our research grid to explore this avalanche of data. Our researchers, experts in the fields of speech recognition, statistical modeling, deep machine learning, and linguistics, use these computational and data resources to continuously advance the boundaries of what can be done with speech technology.

Current applications for consumers and companies.
We optimize our technology for four main application scenarios.

– Our personal assistant solutions enable people to communicate with their devices on human terms. Our systems understand people’s intentions and provide appropriate responses. Drivers operate their GPSs, make phone calls and listen to messages using our robust speech solutions. Speech makes these interactions easier and safer. In that sense, speech technology saves lives. We continuously improve our accuracy, latency, and robustness; and extend our models to new domains, accents, languages, and devices.

– Our document creation solution powers Dragon NaturallySpeaking, the Nuance flagship speech recognition product. We develop a highly personalized speech recognition solution for each user without explicit training. Our solution not only transcribes accurately the words people dictate, but also formats the resulting written documents.

– Medical professionals use our dictation solutions to generate millions of reports every day. We offer both “front-end” solutions, where doctors see and correct reports as they dictate, and “back-end” solutions, where users speak into a microphone, and are later presented with corrected, formatted reports for signature.

– Most spoken communication takes place between people. Our transcription solution accurately converts the spoken words in conversational speech, particularly voicemails, into text. We focus our research on particular challenges in conversational speech, e.g. sloppy formulation and articulation, difficult recording conditions, multiple speakers, and unpredictable content.

Our solutions are implemented in server-based systems, embedded systems, and hybrid systems that use both server and embedded components. We work closely with our hosted operations and frequently roll out new algorithms and models

Where we’re headed next
Here are some representative examples of the problems we research:

  • Acoustic modeling 
    Neural Nets, and particularly “Deep” Neural Nets, provide substantial performance improvements for many speech recognition tasks. Our work around NNs covers network architectures (e.g. DNN, CNN, RNN/LSTM), input features, training algorithms (parallelization, sequence training), and runtime optimizations (and of course other issues). DNNs require well-labeled data and are hard to adapt to new speakers, devices, and acoustic conditions. So we are also interested in using our large corpora of unlabeled data in many languages for DNN training, and in rapidly adapting DNNs for new acoustic conditions.
  • Language modeling
    Recent years have witnessed a loosening of the death-grip of Kneser-Ney (KN) NGram models on state of the art language modeling, with exponential class models (aka Model M) and more recently various large-scale continuous space language models (aka Feed-Forward NN, RNN, LSTM) achieving superior perplexity and word error rate performance over a range of tasks. The recurrent version of these neural models drop the long-held NGram Markov approximation altogether. The gains in performance with these “new” models come with the cost of a significant increase in training times as compared to KN models. Meanwhile cloud-based dictation services have opened the floodgates of (unsupervised) in-domain training data, which KN models are only too happy to consume and benefit from. This presents an interesting challenge: what decisions about model architecture, training implementation and infrastructure (e.g. CPU, GPU, CPU cluster, multiple GPUs), objective function, parameter initialization, optimization method, data selection, and model combination lead to the best performing model within a practical timeframe robustly across application domains? When theory and engineering collide, the most interesting problems are born. We’re having fun solving these problems every day (but don’t worry, there are still a few left).
  • Research engineering
    The research engineering department’s goal is to turn complex algorithms into efficient, robust software. We create and maintain well-engineered toolkits that support researchers’ flexibility to create and test new algorithms, and we create engines and models which support products and applications in over 40 languages. Our customers range from a single researcher trying something new, to millions of users running Nuance products on their own devices or recognition services provided by our data centers. Our model training toolkit runs on our dedicated large-scale computing grid, and our engines run on anything from the smallest devices to large cloud servers.

Explore recent publications by Nuance Speech Recognition researchers.



Selected articles

Investigation on cross- and multilingual mlp features under matched and mismatched acoustical conditions

In this paper, Multi Layer Perceptron (MLP) based multilingual bottleneckfeatures are investigated for acoustic modeling in three languages— German, French, and US English. We use

Read more

Train&align: A new online tool for automatic phonetic alignment

Several automatic phonetic alignment tools have been proposed in the literature. They usually rely on pre-trained speaker-independent models to align new corpora. Their drawback is

Read more

Impact of word error rate on driving performance while dictating short texts

This paper describes the impact of speech recognition word error rate (WER) on driver’s distraction in the context of short message dictation. A multi-modal dictation

Read more

Comparison of grapheme-to-phoneme methods on large pronunciation dictionaries and lvcsr tasks

Grapheme-to-Phoneme conversion (G2P) is usually used within every state-of-the-art ASR system to generalize beyond a fixed set of words. Although the performance is typically already

Read more

Large scale hierarchical neural network language models

Feed-forward neural network language models (NNLMs) are known to improve both perplexity and word error rate performance for speech recognition compared with conventional ngram language

Read more

A hierarchical Bayesian approach for semi-supervised discriminative language modeling

Discriminative language modeling provides a mechanism for differentiating between competing word hypotheses, which are usually ignored in traditional maximum likelihood estimation of N-gram language models.

Read more

Likability Classification – A Not So Deep Neural Network Approach

This papers presents results on the application of restricted Boltzmann machines (RBM) and deep belief networks(DBN) on the Likability Sub-Challenge of the Interspeech 2012 Speaker

Read more

Dictating and editing short texts while driving

Although several existing in-car systems support dictation, there is none which would systematically address dictation and error correction for automotive environments. Dictation and correction systems

Read more

Getting more from morphology in multilingual dependency parsing

We propose a linguistically motivated set of features to capture morphological agreement and add them to the MSTParser dependency parser. Compared to the built-in morphological

Read more

Application of SVM-based correctness predictions to unsupervised discriminative speaker adaptation

The effectiveness of unsupervised speaker adaptation is typicallylimited by errors in the estimated transcription of theadaptation data. Previous work has mitigated this negativeeffect by using

Read more

1 2 3 4 5 18

Upcoming events

See all Research events