Since the birth of Siri, voice interaction has gone through more than ten years. Early voice assistants could only execute simple commands - setting alarms, checking weather, making phone calls. Today, AI speech systems based on big language models are completely breaking down communication barriers between humans and machines, entering an unprecedented new stage.

The most significant progress currently lies in the accuracy and naturalness of speech comprehension. The word error rate of OpenAI's Whisper model has been reduced to below 5% in multiple languages and accents, approaching or even exceeding the human level. Combining the semantic understanding capabilities of large models such as GPT-4, modern voice assistants are no longer just "hearing" user commands, but truly "understanding" user intentions. When you say 'I'm a bit cold', the AI no longer cannot respond, but will actively suggest adjusting the air conditioning temperature or recommend adding a jacket.

Emotional computing is the next frontier in voice interaction. Professor Rosalind Picard from MIT Media Lab proposed this concept as early as 1997, but it was not until the maturity of deep learning technology in recent years that sentiment computing truly began to take root. By analyzing acoustic features such as rhythm, pitch, speed, and breathing patterns in speech, AI systems can recognize users' emotional states in real-time - calm, excited, anxious, angry, or sad. Companies such as Sonic Labs and Hume AI have launched voice APIs that can recognize over 20 emotional states.

At the business application level, AI voice interaction is creating value in multiple industries. The customer service field is the earliest and most mature scenario - intelligent voice customer service can handle more than 80% of routine inquiries, significantly reducing the labor costs of enterprises. In the medical field, AI voice assistants are assisting doctors in completing medical record entry, collecting patient information through natural dialogue, allowing doctors to focus more on diagnosis and treatment itself. The in car voice system is also undergoing a qualitative change - from the frustrating "not understanding" in the past to the current multi round natural conversation, where drivers can interact with the in car AI like chatting with the passenger.

A noteworthy new trend is' personalized voice '. The speech cloning technology launched by companies such as ElevenLabs can generate high fidelity synthesized speech with just a few seconds of speech samples, and users can choose their favorite tone to interact with AI. This personalized experience has transformed voice interaction from a "tool" to a "companion".

However, voice interaction also faces privacy and ethical challenges. The constantly online voice monitoring has raised concerns about data security, and emotion recognition technology may be used to manipulate user emotions. Regulatory agencies in various countries are formulating relevant regulations requiring voice AI devices to obtain explicit consent from users when collecting and processing voice data.

In the future, with the integration of multimodal AI, voice interaction will be combined with sensory channels such as vision and touch to create a more natural immersive interactive experience. The next decade of human-computer interaction will shift from "seeing" to "hearing", and from "manipulation" to "empathy".

This article is a comprehensive compilation of publicly available research and industry reports on speech technology. <｜end▁of▁thinking｜>

Tag: human-computer interaction

New Era of AI Voice Interaction: From Voice Assistants to Emotional Computing

Tag: human-computer interaction

New Era of AI Voice Interaction: From Voice Assistants to Emotional Computing