Jonathan Lalou's Blog

Lecturer

Sébastien Bratières has developed voice-enabled products across Europe since 2001, spanning telephony at Tellme, embedded systems at Voice-Insight, and chat-based dialogue at As An Angel. He currently leads Quint, the voice division of Dawin GmbH. Holding degrees from École Centrale Paris and an MPhil in Speech Processing from the University of Cambridge, he remains active in machine learning research at Cambridge.

Abstract

Sébastien Bratières surveys the landscape of speech recognition technologies available to web developers, contrasting cloud-based APIs with embedded solutions. He covers foundational concepts—acoustic models, language models, grammar-based versus dictation recognition—while evaluating practical trade-offs in latency, accuracy, and deployment. The presentation compares CMU Sphinx, Google Web Speech API, Nuance Developer Network, and Windows Phone 8 Speech API, addressing error handling, dialogue management, and offline capabilities. Developers gain a roadmap for integrating voice into web applications, from rapid prototyping to production-grade systems.

Core Concepts in Speech Recognition: Models, Architectures, and Trade-offs

Bratières introduces the speech recognition pipeline: audio capture, feature extraction, acoustic modeling, language modeling, and decoding. Acoustic models map sound to phonemes; language models predict word sequences.

Grammar-based recognition constrains input to predefined phrases, yielding high accuracy and low latency. Dictation mode supports free-form speech but demands larger models and increases error rates.

Cloud architectures offload processing to remote servers, reducing client footprint but introducing network latency. Embedded solutions run locally, enabling offline use at the cost of computational resources.

Google Web Speech API: Browser-Native Recognition in Chrome

Available in Chrome 25+ beta, the Web Speech API exposes speech recognition via JavaScript. Bratières demonstrates:

const recognition = new webkitSpeechRecognition();
recognition.lang = 'fr-FR';
recognition.onresult = event => console.log(event.results[0][0].transcript);
recognition.start();

Strengths include ease of integration, continuous updates, and multilingual support. Limitations: Chrome-only, requires internet, and lacks fine-grained control over models.

CMU Sphinx: Open-Source Flexibility for Custom Deployments

CMU Sphinx offers fully customizable, embeddable recognition. PocketSphinx runs on resource-constrained devices; Sphinx4 targets server-side Java applications.

Bratières highlights model training: adapt acoustic models to specific domains or accents. Grammar files (JSGF) define valid utterances, enabling precise command-and-control interfaces.

Deployment options span browser via WebAssembly, mobile via native libraries, and server-side processing. Accuracy rivals commercial solutions with sufficient training data.

Nuance Developer Network and Windows Phone 8 Speech API: Enterprise-Grade Alternatives

Nuance provides cloud and embedded SDKs with industry-leading accuracy, particularly in noisy environments. The developer network offers free tiers for prototyping, scaling to paid plans.

Windows Phone 8 integrates speech via the SpeechRecognizerUI class, supporting grammar-based and dictation modes. Bratières notes seamless integration with Cortana but platform lock-in.

Practical Considerations: Latency, Error Handling, and Dialogue Management

Latency varies: cloud APIs achieve sub-second results under good network conditions; embedded systems add processing delays. Bratières advocates progressive enhancement—fallback to text input on failure.

Error handling strategies include confidence scores, n-best lists, and confirmation prompts. Dialogue systems use finite-state machines or statistical models to maintain context.

Embedded and Offline Challenges: Current State and Future Outlook

Bratières addresses offline recognition demand, citing truck drivers embedding systems for navigation. Commercial embedded solutions exist but remain costly.

Open-source alternatives lag in accuracy, particularly for dictation. He predicts convergence: WebAssembly may bring Sphinx-class recognition to browsers, while edge computing reduces cloud dependency.

Conclusion: Choosing the Right Speech Stack

Bratières concludes that no universal solution exists. Prototype with Google Web Speech API for speed; transition to CMU Sphinx or Nuance for customization or offline needs. Voice enables natural interfaces, but success hinges on managing expectations around accuracy and latency.

Links:

• Video of the lecture