Bean Bun APP Updates to 7.2.0, Launches Real-time Voice Models

Recently, the DouBao app was updated to version 7.2.0, introducing a new feature called "Real-Time Voice Mega Model." This model integrates Automatic Speech Recognition (ASR), Natural Language Processing (NLP), and Text-to-Speech (TTS) technologies to offer an all-in-one end-to-end voice conversation solution. Compared to traditional segmented processing methods, this model improves expressiveness, emotional delivery, and response speed, allowing users to interrupt conversations at any time.

Feedback from external users indicates that when compared with GPT-4o, the new model has received higher ratings for voice naturalness and emotional expression. However, specific satisfaction scores and detailed evaluation results have not been disclosed.

Technically, the Real-Time Voice Mega Model undergoes pre-training using multimodal data and is further optimized through reinforcement learning algorithms to enhance model safety and dialogue quality. The development team focused on improving emotional understanding and expression capabilities, aiming to achieve natural conversation while maintaining high intelligence.

Additionally, the model features real-time internet connectivity, enabling it to query and respond with the latest information. For safety considerations, various measures have been implemented to filter out potentially unsafe content, ensuring that both voice and text outputs meet safety standards.

In terms of interaction characteristics, the model delivers low-latency and smooth conversational experiences by optimizing delays in voice generation and comprehension processes. It also takes into account appropriate tones for different scenarios.

Test results show that the new model has received positive evaluations in terms of human-likeness, usefulness, and emotional intelligence, particularly demonstrating progress in capturing and responding to user emotional cues. However, specific data and percentages were not provided as references.