What is StepAudio 2.5 Realtime?
StepAudio 2.5 Realtime is an end-to-end realtime speech model from StepFun. Instead of treating speech recognition, text generation and text-to-speech as completely separate application stages, it is designed as one system for bidirectional spoken interaction. A client can send audio or text and receive streamed audio and text responses.
The model is intended for situations in which latency and conversational flow are important. A voice assistant, for example, can receive a person’s speech through a live connection, interpret both the words and aspects of how they were spoken, and begin returning a spoken response without waiting for a complete conventional text workflow to finish.
The documented model identifier is stepaudio-2.5-realtime. StepFun’s materials describe it as supporting customizable personas, expressive speech and paralinguistic perception. Paralinguistic perception means paying attention to vocal signals beyond the literal words, including hesitation, laughter, sighs, pacing and intonation.
Core capabilities and conversational behavior
StepAudio 2.5 Realtime’s main distinction is its focus on spoken interaction rather than general-purpose language-model tasks. Its documented capabilities include:
- Bidirectional realtime conversation: the system is designed to receive and return media through an ongoing session.
- Paralinguistic understanding: it can interpret vocal cues such as hesitation, laughter, sighs, changes in pace and intonation.
- Expressive speech generation: generated speech can include vocal details such as laughter and sighs, rather than sounding limited to neutral narration.
- Persona customization: developers can configure personality traits, catchphrases, conversational style and emotional boundaries.
- Voice cloning: StepFun documents a reference-audio workflow that produces a custom voice identifier for use with a voice.
These features make the model more relevant to an interactive character or voice companion than to an application that only needs text completion. They may also help when an assistant must respond appropriately to how something was said, not merely to the words transcribed from the audio.
Inputs, outputs and WebSocket integration
The model accepts audio and text inputs and can return audio and text outputs. The official integration uses a bidirectional WebSocket endpoint at wss://api.stepfun.com/v1/realtime?model=stepaudio-2.5-realtime. WebSocket is a persistent connection format that allows the client and server to exchange events while a conversation is in progress, rather than making a separate request for every turn.
Audio can be sent with events such as input_audio_buffer.append. Generated speech is delivered incrementally through streaming events such as response.audio.delta. This event-based design is important for low-latency applications because the client can begin playing response audio while the complete answer is still being generated.
Session configuration supports text and audio modalities, system instructions, a selected voice, PCM16 input and output formats, and server-side voice activity detection. Voice activity detection, or VAD, helps the service identify when a person has started or stopped speaking. The supplied documentation does not publish a production context-window size or maximum output-token limit for this exact endpoint, so those values should be treated as unknown rather than assumed from another StepAudio model.
Position in StepFun’s current lineup
StepAudio 2.5 Realtime belongs to StepFun’s StepAudio family and is specifically the realtime speech-oriented member described in the supplied materials. Its launch announcement presented the model as available, but StepFun’s later public catalog emphasizes StepAudio 3 Realtime. That makes current account-level availability an important deployment check: the documentation and model identifier remain useful references, but access, quotas and commercial terms should be confirmed in the developer account before building a production system.
StepAudio 2.5 Realtime should therefore be viewed as a specialized realtime voice option, not as a replacement for every model in StepFun’s catalog. Its value comes from spoken interaction, streaming behavior and expressive voice features. Applications centered on long-form text, image understanding, image generation or general software development may need a different type of model.
Pricing and availability
StepFun’s Chinese developer documentation lists the following token pricing for StepAudio 2.5 Realtime:
| Usage type | Listed price |
|---|---|
| Uncached input | 10 CNY per 1 million tokens |
| Cached input | 2 CNY per 1 million tokens |
| Output | 70 CNY per 1 million tokens |
The input rates distinguish between new input and cached input, while the output rate is substantially higher than the uncached input rate. Actual voice-application cost will depend on how the service accounts for audio and text tokens, conversation length, repeated instructions and generated speech. Developers should verify the current billing rules and whether these prices remain active for their account.
The model was announced on May 8, 2026, and the launch announcement described it as fully available at that time. Because the later public catalog emphasizes StepAudio 3 Realtime, availability should not be inferred solely from the launch announcement.
Strengths and trade-offs
The model’s strongest fit is a low-latency voice experience. Streaming audio responses, server-side VAD and a persistent WebSocket session can reduce the friction of turn-taking compared with a design that waits for a full response before playback. Expressive output and paralinguistic awareness may also make conversations feel more natural in applications where tone and timing matter.
There are trade-offs. The documented output price is higher than the uncached input price, so long conversations with substantial generated speech may require careful cost controls. The supplied materials do not establish a context limit, maximum output length, general-purpose tool calling, fine-tuning, batch processing or structured-output support for this exact realtime model. Those omissions matter for applications that require strict machine-readable responses, extensive document context or broad workflow automation.
Editorial assessments supplied with the model data rate its realtime speed highly and its coding suitability low. These are comparative editorial scores, not StepFun-published benchmarks. They reflect the model’s positioning as a fast speech-to-speech system rather than evidence that it is a strong programming or reasoning model.
Reasoning, coding and tool support
StepAudio 2.5 Realtime can maintain conversational context and follow instructions within a voice session, but the supplied research does not provide a reasoning benchmark or establish a specialized reasoning capability. It should not be selected for demanding mathematical reasoning, long technical analysis or complex planning without task-specific testing.
The same caution applies to coding. The model can accept text and may be able to discuss technical topics conversationally, but no coding benchmark or coding-specific feature is documented for this endpoint. A conventional text model may be more appropriate for code generation, repository analysis, structured debugging and long source files.
Tool or function calling is not verified in the supplied documentation. Developers should not assume that the realtime event protocol provides general-purpose tools, structured outputs or reliable function invocation. If an application needs those features, they should be confirmed in the current API documentation or implemented through a separately tested orchestration layer.
Best use cases
StepAudio 2.5 Realtime is a good candidate for:
- Voice companions that need natural turn-taking and expressive speech.
- Character roleplay with configurable personality, catchphrases and emotional boundaries.
- Realtime customer-service or task assistants where spoken interaction is the primary interface.
- In-car dialogue systems and other hands-free applications.
- Interview practice, language interaction or coaching experiences that benefit from spoken responses.
- Voice agents that need to react to laughter, hesitation, pacing or other vocal signals.
Voice cloning can be useful when an application needs a consistent custom voice, although consent, rights management and applicable safety requirements should be handled separately by the implementer.
When to choose StepAudio 2.5 Realtime
Choose this model when the central product requirement is a live spoken conversation and the application benefits from streamed audio, expressive delivery or sensitivity to vocal behavior. It is especially attractive when a developer wants one documented realtime speech interface rather than assembling separate speech recognition, language and speech-synthesis components.
Choose another type of model when the primary requirement is image or video generation, embeddings, music generation, large-scale batch text processing, advanced coding or verified structured tool execution. A text-focused model is generally easier to control for long documents and machine-readable output, while a specialized vision or media model is more appropriate for non-audio media tasks. Within StepFun’s own public lineup, the later emphasis on StepAudio 3 Realtime is also a reason to compare current availability and capabilities before committing to StepAudio 2.5 Realtime.
Bottom line
StepAudio 2.5 Realtime is a specialized StepFun model for low-latency, two-way voice interaction. Its documented strengths are streamed audio and text, paralinguistic understanding, expressive speech, persona customization and a WebSocket interface with server-side VAD. Its main uncertainties are equally important: no published context or maximum-output limit for this endpoint, and no verified general tool calling, fine-tuning, batch or structured-output support. It is best evaluated as a realtime voice engine, not as a general-purpose replacement for text, coding or media models.