Detección de pronunciación para la nueva experiencia de aprendizaje de inglés de Alexa

El aumento de datos, novedosas funciones de pérdidas y un entrenamiento con poca supervisión permiten crear un modelo innovador para detectar errores de pronunciación.

En enero de 2023, Alexa lanzó en España una experiencia de aprendizaje de idiomas para ayudar a los hispanohablantes a aprender inglés para principiantes. Esta experiencia se desarrolló en colaboración con Vaughan, el principal proveedor de aprendizaje de inglés en España, con el objetivo de ofrecer un programa de aprendizaje de inglés inmersivo centrado en la mejora de la pronunciación.

Ahora estamos ampliando esta oferta a México y a la población de habla hispana de Estados Unidos, y en el futuro planeamos añadir más idiomas. Esta experiencia de aprendizaje de idiomas incluye lecciones estructuradas de vocabulario, gramática, expresión y pronunciación, con ejercicios prácticos y pruebas. Para probarla, configura el idioma de tu dispositivo a español y dile a Alexa "Quiero aprender inglés".

Mini-lesson content page_ES.png
Página de contenidos de lecciones cortas: lecciones de vocabulario, gramática, expresión y pronunciación.

Lo más destacado de esta skill de Alexa es su función de pronunciación, la cual proporciona información precisa cada vez que un cliente pronuncia mal una palabra o frase. En la Conferencia Internacional de Acústica, Habla y Procesamiento de Señales (ICASSP por sus siglas en inglés) de este año, presentamos un artículo en el que describíamos nuestro innovador método de detección de errores de pronunciación.

alexaspeechspanish.jpg
Corrección de pronunciación: El texto en azul indica una pronunciación correcta, mientras que el rojo indica una pronunciación incorrecta. Para frases/palabras pronunciadas incorrectamente, Alexa brindará instrucciones detalladas sobre cómo pronunciarlas.

Nuestro método utiliza un novedoso modelo fonético de redes neuronales recurrentes (RNN-T por sus siglas en inglés) que predice los fonemas, las unidades más pequeñas del habla, a partir de la pronunciación del alumno. Por lo tanto, el modelo puede proporcionar una evaluación detallada de la pronunciación a nivel de palabra, sílaba o fonema. Por ejemplo, si un alumno pronuncia incorrectamente la palabra "rabbit" como "rabid", el modelo mostrará la secuencia de cinco fonemas R AE B IH D. Posteriormente, puede detectar los fonemas (IH D) y la sílaba (-bid) mal pronunciados utilizando la alineación de Levenshtein para comparar la secuencia de fonemas con la secuencia de referencia "R AE B AH T".

El artículo destaca dos brechas de conocimiento que no se habían abordado en anteriores modelos de pronunciación. La primera es la capacidad de distinguir fonemas similares en distintos idiomas (por ejemplo, la "r" rodada en español vs. la "r" en inglés). Para ello, diseñamos un léxico de pronunciación multilingüe y creamos un inmenso conjunto de datos fonéticos mixtos para el programa de aprendizaje.

La otra brecha de conocimiento es la capacidad de aprender patrones únicos de pronunciación errónea de los alumnos de idiomas. Para ello, aprovechamos la autorregresividad del modelo RNN-T, es decir, la dependencia de sus resultados de las entradas y salidas anteriores. Este conocimiento del contexto significa que el modelo puede captar patrones frecuentes de pronunciación errónea a partir de los datos del entrenamiento. Nuestro modelo de pronunciación ha obtenido los mejores resultados tanto en precisión de predicción de fonemas, como de detección de errores de pronunciación.

Aumento de datos L2

Uno de los principales retos técnicos a la hora de crear un modelo de reconocimiento fonético para hablantes no nativos (L2) es que los conjuntos de datos para el diagnóstico de errores de pronunciación son muy limitados. En nuestro artículo de Interspeech 2022 "L2-GEN: Un enfoque neuronal de parafraseo de fonemas para el diagnóstico de errores de pronunciación en síntesis del habla L2", planteamos cerrar esta brecha mediante el incremento de datos. En concreto, creamos un parafraseador de fonemas que puede generar fonemas realistas de L2 para hablantes de un lugar específico, por ejemplo, fonemas que representen a un hablante nativo de español hablando en inglés.

Como es habitual en las tareas de corrección de errores gramaticales, utilizamos un modelo de secuencia a secuencia, pero invertimos la dirección de la tarea entrenando al modelo para pronunciar mal las palabras en lugar de corregir los errores de pronunciación. Además, para enriquecer y diversificar aún más las secuencias de fonemas L2 generados, propusimos un componente de decodificación diversificado y consciente de las preferencias que combina una búsqueda en haz diversificada con una pérdida de preferencia que se inclina hacia los errores de pronunciación similares a los humanos.

Para cada tono de entrada o fragmento del habla, el modelo produce varios fonemas posibles como salidas, y las secuencias de fonemas se modelan como un árbol en el que las posibilidades proliferan con cada nuevo tono. Normalmente, las secuencias de fonemas mejor clasificadas se extraen del árbol mediante las búsquedas en haz que persigue solo las ramas del árbol con las probabilidades más altas. En nuestro trabajo, sin embargo, propusimos un método de búsqueda en haz que da prioridad a los fonemas inusuales, o candidatos a fonema que difieren de la mayoría de los demás en la misma profundidad del árbol.

A partir de fuentes establecidas en la documentación sobre aprendizaje de idiomas, también elaboramos listas de errores de pronunciación comunes a nivel de fonema, representados como pares de fonemas, uno del fonema estándar de la lengua y otro de su variante no estándar. Construimos una función de pérdida que, durante el proceso de aprendizaje del modelo, da prioridad a los resultados que utilizan las variantes no estándar de nuestra lista.

En los experimentos observamos mejoras de precisión de hasta el 5% en la detección de errores de pronunciación con respecto a un modelo de referencia entrenado sin datos adicionales.

Equilibrando el falso rechazo y la falsa aceptación

Una consideración clave a la hora de diseñar un modelo de pronunciación para una experiencia de aprendizaje de idiomas es equilibrar la proporción de falsos rechazos y falsas aceptaciones. Un falso rechazo se produce cuando el modelo de pronunciación detecta un error de pronunciación, pero en realidad el alumno estaba en lo cierto o utilizaba una pronunciación coherente pero ligeramente acentuada. Una falsa aceptación se produce cuando un alumno pronuncia mal una palabra y el modelo no lo detecta.

Nuestro sistema tiene dos características de diseño enfocadas a equilibrar estas dos métricas. Para reducir las falsas aceptaciones, primero combinamos nuestros léxicos de pronunciación estándar para inglés y español en un léxico único con múltiples fonemas correspondientes a cada palabra. Después, utilizamos ese léxico para analizar automáticamente muestras de habla no comentadas que se clasifican en tres categorías: español nativo, inglés nativo y español e inglés codificados. El entrenamiento del modelo con este conjunto de datos le permite distinguir diferencias muy sutiles entre fonemas.

Para reducir los falsos rechazos utilizamos un léxico de pronunciación multirreferencial en el que cada palabra se asocia a varias pronunciaciones de referencia. Por ejemplo, la palabra "data" puede pronunciarse como "day-tah" o "dah-tah" y el sistema aceptará ambas variaciones como correctas.

Actualmente seguimos estudiando varios métodos para mejorar nuestra función de evaluación de la pronunciación. Uno de ellos es la creación de un modelo multilingüe que pueda utilizarse para evaluar la pronunciación en muchos idiomas. También estamos ampliando el modelo para diagnosticar más características de pronunciación errónea, como el tono y el acento léxico.

Research areas

Related content

US, NY, New York
We are seeking an Applied Scientist to develop and optimize Visual Inertial Odometry (VIO) and sensor fusion systems for our intelligent robots. In this role, you will design, implement, and deploy state estimation and tracking algorithms that enable robots to understand their position and motion in real time, even in challenging and dynamic environments. You will own the full pipeline from algorithm development through embedded deployment, ensuring that perception systems run efficiently on resource-constrained robotic hardware. You will also leverage modern machine learning approaches to push the boundaries of classical perception methods, combining learned representations with geometric techniques to achieve robust, real-time performance. This is a deeply hands-on role. You will work directly with sensors, hardware, and real-world data, while prototyping, testing, and iterating in physical environments. The ideal candidate has strong foundations in VIO and sensor fusion, practical experience optimizing algorithms for embedded platforms, and familiarity with how modern deep learning is transforming perception. Key job responsibilities - Design and implement Visual Inertial Odometry algorithms for robust real-time state estimation on robotic platforms like Sprout - Develop multi-sensor fusion pipelines integrating cameras, IMUs, and other sensing modalities for accurate pose tracking - Optimize perception and tracking algorithms for deployment on embedded hardware (e.g., ARM, GPU-accelerated edge devices) under strict latency and power constraints - Apply modern ML-based perception techniques (learned features, depth estimation, neural odometry) to complement and improve classical geometric approaches - Build and maintain calibration, evaluation, and benchmarking infrastructure for perception systems - Collaborate with hardware, controls, and navigation teams to integrate perception outputs into the robot’s autonomy stack - Lead technical projects from research prototyping through production deployment
US, WA, Seattle
Innovators wanted! Are you an entrepreneur? A builder? A dreamer? This role is part of an Amazon Special Projects team that takes the company’s Think Big leadership principle to the limits. If you’re interested in innovating at scale to address big challenges in the world, this is the team for you. As an Applied Scientist on our team, you will focus on building state-of-the-art ML models for biology. Our team rewards curiosity while maintaining a laser-focus in bringing products to market. Competitive candidates are responsive, flexible, and able to succeed within an open, collaborative, entrepreneurial, startup-like environment. At the forefront of both academic and applied research in this product area, you have the opportunity to work together with a diverse and talented team of scientists, engineers, and product managers and collaborate with other teams. Key job responsibilities - Build, adapt and evaluate ML models for life sciences applications - Collaborate with a cross-functional team of ML scientists, biologists, software engineers and product managers
US, MA, Boston
MULTIPLE POSITIONS AVAILABLE Employer: AMAZON.COM SERVICES LLC Offered Position: Economist III Job Location: Boston, Massachusetts Job Number: AMZ9898444 Position Responsibilities: Mentor and guide the applied scientists and economists in our organization and hold us to a high standard of technical rigor and excellence in science. Design and lead roadmaps for complex science projects to help SP have a delightful selling experience while creating long term value for our shoppers. Work with our engineering partners and draw upon your experience to meet latency and other system constraints. Identify untapped, high-risk technical and scientific directions, and simulate new research directions that you will drive to completion and deliver. Be responsible for communicating our science innovations to the broader internal & external scientific community. Position Requirements: Ph.D. or foreign equivalent degree in Economics or a related field and two years of research or work experience in the job offered or a related occupation. Must have two years of research or work experience in the following skill(s): 1) experience in econometrics including experience with program evaluation, forecasting, time series, panel data, or high dimensional problems; 2) experience with economic theory and quantitative methods; and 3) coding in a scripting language such as R, Python, or similar. Amazon.com is an Equal Opportunity-Affirmative Action Employer – Minority / Female / Disability / Veteran / Gender Identity / Sexual Orientation. 40 hours / week, 8:00am-5:00pm, Salary Range $159,200/year to $215,300/year. Amazon is a total compensation company. Dependent on the position offered, equity, sign-on payments, and other forms of compensation may be provided as part of a total compensation package, in addition to a full range of medical, financial, and/or other benefits. For more information, visit: https://www.aboutamazon.com/workplace/employee-benefits.#0000
US, WA, Seattle
Applied Scientists in AWS Automated Reasoning are dedicated to making AWS the best computing service in the world for customers who require advanced and rigorous solutions for automated reasoning, privacy, and sovereignty. Key job responsibilities - Solve large or significantly complex problems that require deep knowledge and understanding of your domain and scientific innovation. - Own strategic problem solving, and take the lead on the design, implementation, and delivery for solutions that have a long-term quantifiable impact. - Provide cross-organizational technical influence, increasing productivity and effectiveness by sharing your deep knowledge and experience. - Develop strategic plans to identify fundamentally new solutions for business problems. - Assist in the career development of others, actively mentoring individuals and the community on advanced technical issues.
US, WA, Seattle
Amazon's Worldwide Pricing & Promotions organization is seeking a talented, hands-on Research Scientist to join the Pricing and Promotion Optimization Science (P2OS) team — the optimization "application layer" within Amazon's Pricing Sciences organization. Amazon adjusts prices on hundreds of millions of products daily across a global marketplace; P2OS is the team that makes those prices optimal. P2OS is a small, specialized unit with an outsized charter: develop and maintain the models that determine optimal prices and promotions across Amazon's catalog and merchant programs. We own the full optimization stack — from price prediction to promotion targeting to competitiveness guardrails — and we measure success in terms of accretive Gross Contribution and Customer Pricing Perception (GCCP). Our work spans Retail Core, Amazon Business, Fresh, Grocery, and international marketplaces, and we are continually investing in more extensible, generalizable science foundations to keep pace with a growing and evolving business. We are looking for an innovative, organized, and customer-focused scientist with exceptional machine learning and predictive modeling skills, causal and experimental evaluation experience, and the entrepreneurial spirit to apply state-of-the-art methods to some of the most impactful pricing problems in e-commerce. You should be comfortable with ambiguity, motivated by measurable business impact, and excited by the opportunity to work at Amazon-scale. Key job responsibilities * Innovate and build. Design, develop, and deploy machine learning models that set optimal prices and promotions across Amazon's global catalog. Own models end-to-end — from problem formulation and data analysis through offline evaluation, A/B testing, and production launch. * Build a generalizable science foundation. Develop models and evaluation frameworks designed to scale across merchant programs, product categories, and marketplaces — enabling cross-learning and reducing the time and cost of applying science to new business contexts. * Build and evolve optimization systems. Design and improve optimization systems — including reinforcement learning and multi-objective optimization approaches — that automate price and promotion decisions at scale across millions of products. * Apply generative AI and foundation models. Identify and pursue opportunities to leverage large language models, embeddings, and generative AI techniques in pricing science — from enriching product representations and extracting competitive signals from unstructured data, to building more capable and explainable pricing systems. * Experiment rigorously. Design and execute A/B tests and causal inference studies to measure the business and customer impact of pricing model changes. Translate findings into production-ready science improvements. * Stay at the frontier. Establish mechanisms to track the latest advances in reinforcement learning, causal ML, multi-objective optimization, generative AI, and demand modeling — and identify opportunities to apply them to Pricing & Promotions business problems. * See the big picture. Contribute to the long-term scientific vision for how Amazon sets competitive, perception-preserving prices — balancing profitability, customer trust, and marketplace health.
US, CA, San Francisco
Amazon is on a mission to redefine the future of automation — and we're looking for exceptional talent to help lead the way. We are building the next generation of advanced robotic systems that seamlessly blend cutting-edge AI, sophisticated control systems, and novel mechanical design to create adaptable, intelligent automation solutions capable of operating safely alongside humans in dynamic, real-world environments. At Amazon, we leverage the power of machine learning, artificial intelligence, and advanced robotics to solve some of the most complex operational challenges at a scale unlike anywhere else in the world. Our fleet of robots spans hundreds of facilities globally, working in sophisticated coordination to deliver on our promise of customer excellence — and we're just getting started. As a Sr. Scientist in Robot Navigation, you will be at the forefront of this transformation — architecting and delivering navigation systems that are intelligent, safe, and scalable. You will bring deep expertise in learning-based planning and control, a strong understanding of foundation models and their application to embodied agents, and as well as have in-depth understanding of control-theoretic approaches such as model predictive control (MPC)-based trajectory planning. You will develop navigation solutions that seamlessly blend data-driven intelligence with principled control-theoretic guarantees. Our vision is bold: to build navigation systems that allow robots to move fluidly and safely through dynamic environments — understanding context, anticipating change, and adapting in real time. You will lead research that bridges the gap between cutting-edge academic advances and production grade deployment, collaborating with world-class teams pushing the boundaries of robotic autonomy, manipulation, and human-robot interaction. Join us in building the next generation of intelligent navigation systems that will define the future of autonomous robotics at scale. Key job responsibilities - Design, develop, and deploy perception algorithms for robotics systems, including object detection, segmentation, tracking, depth estimation, and scene understanding - Lead research initiatives in computer vision, sensor fusion and 3D perception - Collaborate with cross-functional teams including robotics engineers, software engineers, and product managers to define and deliver perception capabilities - Drive end-to-end ownership of ML models — from data collection and labeling strategy to training, evaluation, and deployment - Mentor junior scientists and engineers; contribute to a culture of technical excellence - Define and track key metrics to measure perception system performance in real-world environments - Publish research findings in top-tier venues (CVPR, ICCV, ECCV, ICRA, NeurIPS, etc.) and contribute to patents A day in the life - Train ML models for deployment in simulation and real-world robots, identify and document their limitations post-deployment - Drive technical discussions within your team and with key stakeholders to develop innovative solutions to address identified limitations - Actively contribute to brainstorming sessions on adjacent topics, bringing fresh perspectives that help peers grow and succeed — and in doing so, build lasting trust across the team - Mentor team members while maintaining significant hands-on contribution to technical solutions About the team Our team is a group is a diverse group of scientists and engineers passionate about building intelligent machines. We value curiosity, rigor, and a bias for action. We believe in learning from failure and iterating quickly toward solutions that matter.
US, NY, New York
The Sponsored Products and Brands team at Amazon Ads is re-imagining the advertising landscape through generative AI technologies, revolutionizing how millions of customers discover products and engage with brands across Amazon.com and beyond. We are at the forefront of re-inventing advertising experiences, bridging human creativity with artificial intelligence to transform every aspect of the advertising lifecycle from ad creation and optimization to performance analysis and customer insights. We are a passionate group of innovators dedicated to developing responsible and intelligent AI technologies that balance the needs of advertisers, enhance the shopping experience, and strengthen the marketplace. If you're energized by solving complex challenges and pushing the boundaries of what's possible with AI, join us in shaping the future of advertising. About the team SPB Agent team's vision is to build a highly personalized and context-aware agentic advertiser guidance system that seamlessly integrates Large Language Models (LLMs) with sophisticated tooling, operating across all experiences. The SPB-Agent is the central agent that interfaces with advertisers across Ads Console, Selling Partner portals (Seller Central, KDP, Vendor Central), and internal Sales systems. We identify high-impact opportunities spanning from strategic product guidance to granular optimization and deliver them through personalized, scalable experiences grounded in state-of-the-art agent architectures, reasoning frameworks, sophisticated tool integration, and model customization approaches including fine-tuning, MCP, and preference optimization. This presents an exceptional opportunity to shape the future of e-commerce advertising through advanced AI technology at unprecedented scale, creating solutions that directly impact millions of advertisers.
US, WA, Seattle
Applied Scientists in AWS Automated Reasoning are dedicated to making AWS the best computing service in the world for customers who require advanced and rigorous solutions for automated reasoning, privacy, and sovereignty. Key job responsibilities The successful candidate will: - Solve large or significantly complex problems that require deep knowledge and understanding of your domain and scientific innovation. - Own strategic problem solving, and take the lead on the design, implementation, and delivery for solutions that have a long-term quantifiable impact. - Provide cross-organizational technical influence, increasing productivity and effectiveness by sharing your deep knowledge and experience. - Develop strategic plans to identify fundamentally new solutions for business problems. - Assist in the career development of others, actively mentoring individuals and the community on advanced technical issues. A day in the life This is a unique and rare opportunity to get in early on a fast-growing segment of AWS and help shape the technology, product and the business. You will have a chance to utilize your deep technical experience within a fast moving, start-up environment and make a large business and customer impact. About the team Diverse Experiences Amazon Automated Reasoning values diverse experiences. Even if you do not meet all of the qualifications and skills listed in the job description, we encourage candidates to apply. If your career is just starting, hasn't followed a traditional path, or includes alternative experiences, don't let it stop you from applying. Why Amazon Automated Reasoning? At Amazon, automated reasoning is central to maintaining customer trust and delivering delightful customer experiences. Our organization is responsible for creating and maintaining a high bar for automated reasoning across all of Amazon's products and services. We offer talented automated reasoning professionals the chance to accelerate their careers with opportunities to build experience in a wide variety of areas including cloud, devices, retail, entertainment, healthcare, operations, and physical stores. Inclusive Team Culture In Amazon Automated Reasoning, it's in our nature to learn and be curious. Ongoing DEI events and learning experiences inspire us to continue learning and to embrace our uniqueness. Addressing the toughest automated reasoning challenges requires that we seek out and celebrate a diversity of ideas, perspectives, and voices. Training & Career Growth We're continuously raising our performance bar as we strive to become Earth's Best Employer. That's why you'll find endless knowledge-sharing, training, and other career-advancing resources here to help you develop into a better-rounded professional. Work/Life Balance We value work-life harmony. Achieving success at work should never come at the expense of sacrifices at home, which is why flexible work hours and arrangements are part of our culture. When we feel supported in the workplace and at home, there's nothing we can't achieve.
US, WA, Seattle
Applied Scientists in AWS Automated Reasoning are dedicated to making AWS the best computing service in the world for customers who require advanced and rigorous solutions for automated reasoning, privacy, and sovereignty. Key job responsibilities The successful candidate will: - Solve large or significantly complex problems that require deep knowledge and understanding of your domain and scientific innovation. - Own strategic problem solving, and take the lead on the design, implementation, and delivery for solutions that have a long-term quantifiable impact. - Provide cross-organizational technical influence, increasing productivity and effectiveness by sharing your deep knowledge and experience. - Develop strategic plans to identify fundamentally new solutions for business problems. - Assist in the career development of others, actively mentoring individuals and the community on advanced technical issues. A day in the life This is a unique and rare opportunity to get in early on a fast-growing segment of AWS and help shape the technology, product and the business. You will have a chance to utilize your deep technical experience within a fast moving, start-up environment and make a large business and customer impact. About the team Diverse Experiences Amazon Automated Reasoning values diverse experiences. Even if you do not meet all of the qualifications and skills listed in the job description, we encourage candidates to apply. If your career is just starting, hasn't followed a traditional path, or includes alternative experiences, don't let it stop you from applying. Why Amazon Automated Reasoning? At Amazon, automated reasoning is central to maintaining customer trust and delivering delightful customer experiences. Our organization is responsible for creating and maintaining a high bar for automated reasoning across all of Amazon's products and services. We offer talented automated reasoning professionals the chance to accelerate their careers with opportunities to build experience in a wide variety of areas including cloud, devices, retail, entertainment, healthcare, operations, and physical stores. Inclusive Team Culture In Amazon Automated Reasoning, it's in our nature to learn and be curious. Ongoing DEI events and learning experiences inspire us to continue learning and to embrace our uniqueness. Addressing the toughest automated reasoning challenges requires that we seek out and celebrate a diversity of ideas, perspectives, and voices. Training & Career Growth We're continuously raising our performance bar as we strive to become Earth's Best Employer. That's why you'll find endless knowledge-sharing, training, and other career-advancing resources here to help you develop into a better-rounded professional. Work/Life Balance We value work-life harmony. Achieving success at work should never come at the expense of sacrifices at home, which is why flexible work hours and arrangements are part of our culture. When we feel supported in the workplace and at home, there's nothing we can't achieve.
GB, London
Are you excited about using econometrics, experimentation, and machine learning to impact real-world business decisions? We are looking for an Economist II to work on challenging problems at the intersection of causal inference and machine learning for Prime Video Ads. You will design experiments, build econometric and ML models, and translate findings into decisions that shape how millions of customers experience advertising on Prime Video. If you have a deeply quantitative approach to problem-solving, enjoy building and implementing models end-to-end, and want to work on problems where rigorous economics meets production-scale ML, we want to talk to you. Key job responsibilities - Design, execute, and analyze experiments to measure the impact of ad policies on customer behavior and business outcomes - Develop causal inference models (experimental and observational) to estimate short- and long-term effects of strategic initiatives - Collaborate with scientists, engineers, and product teams to deliver measurable business impact - Influence business leaders based on empirical findings