StepAudio 2.5 Realtime - A real-time speech large model launched by StepFun

What is StepAudio 2.5 Realtime

StepAudio 2.5 Realtime is an end-to-end real-time speech large model launched by StepFun, focusing on a human-level voice conversation experience. The model supports deep interaction at the content level and is completely close to real humans in voice expressiveness. It features three core breakthroughs: top-tier paralinguistic capabilities, tens of millions of customizable personas, and leading dual intelligence in conversation, creating an AI chat companion that is warm, soulful, and has attitude.

Main features of StepAudio 2.5 Realtime

• Top-tier paralinguistic perception: Accurately captures intonation, speaking rate, pauses, and even sighs and chuckles, understanding the subtext and emotional shifts in conversations.
• Tens of millions of customizable personas: From personality traits and background experiences to language habits and conversation boundaries, it supports fine-grained adjustment across all dimensions to create a unique exclusive character.
• Leading dual intelligence in conversation: Deeply understands complex semantics and cleverly throws in memes, while also possessing high emotional intelligence feedback capabilities, enabling in-depth and insightful communication.
• Real-time voice interaction: End-to-end real-time conversation architecture, supports Chinese and English, with fast and natural fluent responses.
• Roleplay stability: Specially optimized for Roleplay scenarios, it can still firmly adhere to the preset persona under extreme stress tests, avoiding persona collapse.

Technical principles of StepAudio 2.5 Realtime

• Million-level persona data augmentation: Based on over 10,000 high-quality native personas, it generates a million-level persona feature matrix through algorithmic fission, and integrates massive real-scenario conversation corpora for training, building an extremely strong data generalization foundation for the model, allowing it to respond steadily even to long-tail topics.
• Roleplay-exclusive RLHF alignment: Deep reinforcement learning alignment optimization for role-playing scenarios, solving the most common OOC (out-of-character) problem in AI role-playing. Under extreme adversarial stress tests, the model can still maintain extremely stable character performance capabilities.
• Deep integration of understanding and generation: Fully inherits StepAudio 2.5 TTS capabilities, and through reinforcement learning deeply couples speech understanding and generation, achieving the dual capabilities of "global scene setting" and "intra-sentence detail refinement", accurately perceiving the conversation atmosphere and responding with a matching voice texture.

How to use StepAudio 2.5 Realtime

• Apply for access: Visit the StepFun Open Platform https://platform.stepfun.com/docs/zh/guides/models/stepaudio-2.5-realtime，注册账号并获取 API key, and developers access the real-time voice service via the WebSocket protocol.
• Configure parameters: After connecting, send the session.update instruction to set the audio format (such as pcm16) and select the model version.
• Customize persona: Define the character's personality, verbal tics, voice timbre, and conversation boundaries in detail in the instruction, achieving freely customizable personas at the tens of millions level.
• Start conversation: After establishing a connection, two-way real-time voice streaming can be started, and the model will automatically perceive emotions and generate responses with paralinguistic details.
• Online experience: Ordinary users do not need code; they can directly visit the StepFun Experience Center, select a preset persona, and start human-like voice small talk.

Key information and usage requirements of StepAudio 2.5 Realtime

• Product name：StepAudio 2.5 Realtime
• Development team: StepFun
• Product positioning: End-to-end real-time speech large model, human-like conversation and full-dimensional persona customization
• Supported languages: Chinese, English
• Usage requirements: Developers need an API key to connect via WebSocket; regular users can try it directly in the experience center on the official website

Core advantages of StepAudio 2.5 Realtime

• Industry-leading paralinguistic perception: Scored 82.18 in paralinguistic understanding tests, with precise perception of acoustic features such as speaking rate, emotion, and age.
• Leading across the board in evaluations: Covers five major dimensions: subjective evaluation, general conversation, in-vehicle scenarios, paralinguistic understanding, and speech question answering, ranking first in all of them.
• Stable persona that does not collapse: Dedicated RLHF alignment optimization ensures character consistency under extreme situations, delivering an immersive experience far beyond similar products.
• Extremely human-like: Scored 80.41 in subjective human evaluations, naturally incorporating real details such as light laughter and sighs, with conversational quality fully comparable to a real friend.

Project address of StepAudio 2.5 Realtime

• Project official website：https://stepaudiollm.github.io/step-audio-2.5-realtime/
• Online experience：https://www.stepfun.com/studio/audio?tab=voice-chat

Comparison of StepAudio 2.5 Realtime with similar competitors

Application scenarios of StepAudio 2.5 Realtime

• Emotional companionship: Bedtime heart-to-heart talks, emotional soothing, and complaint-style interaction, providing real-friend-like companionship with maximum empathy.
• Role-playing: Freely customize any persona, from sweet girls to domineering CEOs, meeting immersive needs such as games, novels, and virtual socializing.
• Knowledge interaction: Quick knowledge Q&A, Feihualing, and brain teasers, with deep understanding and engaging interaction capabilities.
• Skills training: High-intensity mock interviews, in-depth follow-up questions, and professional-level feedback, with interview training depth far beyond similar products.
• In-vehicle assistant: It remains stable and smooth even in noisy environments, supporting natural interaction and task completion for navigation, vehicle control, information queries, and more.