What is StepAudio 2.5 TTS?
StepAudio 2.5 TTS is a specialized text-to-speech model provided by StepFun. Its purpose is to convert written text into spoken audio while responding to natural-language directions about how that text should be performed.
Rather than limiting users to a small set of fixed voices or basic emotion labels, the model accepts contextual instructions such as speaking in a restrained tone, sounding hurried, lowering the voice, adding a short breath, or performing a line with sadness. This makes it particularly relevant to narration, character dialogue, voiceovers, and conversational applications where delivery matters as much as the words themselves.
The model is available through StepFun’s Open Platform. StepFun’s current catalog also includes StepAudio 3 TTS, but StepAudio 2.5 TTS remains a documented model with its own identifier and pricing. The documented model ID is stepaudio-2.5-tts.
How its context control works
StepAudio 2.5 TTS provides two levels of performance control. A global instruction describes the general way the complete passage should sound. It can establish the scene, emotional state, vocal style, or overall delivery.
Inline directions are inserted into the text in parentheses. These directions can alter a specific moment without changing the instruction for the entire passage. For example, a script could include directions for a short pause, a breath, stronger emphasis, a sudden change in emotion, or a quieter delivery. The directions are treated as synthesis instructions rather than words to be spoken.
This approach is useful when a single passage contains several performance changes. A global instruction might establish a tense nighttime conversation, while inline directions can specify where a character hesitates, whispers, or reacts with surprise. The supplied documentation describes this as natural-language control, so applications do not need to rely exclusively on a proprietary emotion-label system.
Zero-shot voice cloning and audio input
The model supports zero-shot voice cloning from a short reference recording. According to StepFun’s documentation, approximately three seconds of reference audio can be used to reproduce a target voice while retaining the model’s global and inline expression controls.
In practical terms, this means the reference recording supplies the vocal identity, while the text and instructions determine what the voice says and how the line is delivered. This combination can be useful for character production, localized content, interactive agents, and other applications that need a consistent voice without training a separate custom model.
Voice cloning requires appropriate authorization. Developers should obtain consent from the speaker, respect applicable rights, and consider disclosure requirements when generated speech could be mistaken for an authentic recording. The capability should not be treated as permission to reproduce any voice merely because a recording is available.
Supported inputs and outputs
StepAudio 2.5 TTS primarily accepts text and produces spoken audio. It can also accept reference audio for voice cloning, which is why its documented input modalities include text and audio. Its output is audio rather than text, images, video, or structured data.
| Capability | Supported detail |
|---|---|
| Text input | Yes |
| Reference audio input | Yes, for zero-shot voice cloning |
| Audio output | Yes, synthesized speech |
| Image or video input | Not documented |
| Image or video output | Not supported as the model’s purpose |
| Streaming | Yes, including low-latency audio generation |
There is no evidence in the supplied documentation that the model supports general image understanding, video processing, transcription, text generation, or multimodal reasoning. Its audio capability is specifically directed toward speech synthesis.
Input limits and API access
Each request supports up to 1,000 input characters. The global instruction is limited to 200 characters. These limits are important for production planning: a long audiobook chapter, script, or article must be divided into multiple requests rather than submitted as one unrestricted passage.
Long-form applications should keep the voice settings and broad delivery instruction consistent across segments. They may also need to manage transitions between requests so that pauses, emotional continuity, and scene changes do not sound abrupt.
StepFun documents two access patterns. Non-streaming synthesis uses the POST /v1/audio/speech endpoint and returns a complete audio file. Streaming synthesis uses a WebSocket endpoint at /v1/realtime/audio, which is intended for lower-latency playback and interactive experiences.
Streaming is especially relevant to conversational agents and interfaces where the listener should begin hearing speech before the full response has been rendered. Non-streaming requests are simpler when an application needs a complete audio asset for storage, editing, or later delivery.
StepAudio 2.5 TTS pricing
StepFun documents contextual text-to-speech pricing at ¥5.8 per 10,000 characters. Voice cloning or voice creation is charged separately at ¥9.9 per voice, according to the model documentation.
The pricing model is character-based rather than token-based. That makes it easier to estimate the cost of a fixed script, but the total depends on the amount of text submitted and whether the application creates or clones additional voices. Audio output is included in the synthesis price under the documented pricing structure.
Because the model has a 1,000-character request limit, applications processing longer material should calculate cost across multiple requests. The supplied research does not provide a separate streaming surcharge, volume discount, currency conversion, or recurring subscription price, so those details should not be assumed.
Main strengths and limitations
Where the model is strong
- Expressive control: Natural-language global and inline instructions allow users to describe performance details such as mood, pacing, pauses, breathing, emphasis, and vocal style.
- Voice consistency: Zero-shot voice cloning can reproduce a target voice from approximately three seconds of reference audio without a separately trained voice model.
- Interactive delivery: WebSocket streaming supports applications that need audio to begin playing with low latency.
- Focused workflow: The model is designed specifically for speech production, so its controls are directed toward vocal performance rather than general-purpose text generation.
Important limitations
- Short request size: The 1,000-character input limit is restrictive for long scripts and requires segmentation.
- Instruction size: The global instruction is limited to 200 characters, so complex scene direction must be expressed concisely or distributed through inline directions.
- Narrow task scope: It is not a general language model for reasoning, coding, research, image generation, or video generation.
- No documented tool support: The supplied specifications do not indicate function calling, external tools, web search, or structured-output generation.
- Voice governance requirements: Cloned voices require authorization and responsible handling of identity and disclosure.
StepFun does not publish directly comparable reasoning, coding, speed, or cost scores for this model. Those editorial score fields are therefore best treated as unverified rather than converted into numerical ratings. The model’s practical speed advantage comes from its documented streaming mode, but no standardized latency benchmark is supplied.
When to choose StepAudio 2.5 TTS
Choose StepAudio 2.5 TTS when the main requirement is expressive speech rather than general AI capability. It is a suitable candidate for:
- Audiobook-style or story narration with controlled emotion and pacing
- Character dialogue for games, animation, and short-form video
- Voiceover production that needs local changes in emphasis or delivery
- Emotionally nuanced virtual agents and conversational interfaces
- Interactive playback that benefits from streamed audio
- Authorized voice cloning from a short reference recording
It is also a reasonable fit when a production team wants to specify vocal behavior in ordinary language instead of building a large collection of fixed voice presets. Inline directions can make a script more explicit about performance without requiring those directions to be spoken.
When another option may be more appropriate
Another speech model may be a better choice if the application requires unrestricted long-form input in a single request, a published latency benchmark, a broader selection of documented languages or voices, or a different pricing structure. The supplied research does not establish those comparative details for competing products, so the choice should be based on verified requirements rather than assumed superiority.
A general-purpose language model is more appropriate for tasks such as reasoning over documents, writing code, calling tools, or producing structured business data. StepAudio 2.5 TTS can turn prepared text into speech, but it is not documented as the system that should create or analyze that content.
For a long production workflow, a separate text-generation or orchestration system may prepare and segment the script, while StepAudio 2.5 TTS handles the final spoken output. That architecture acknowledges the model’s narrow purpose and its per-request character limit instead of treating it as an all-purpose assistant.
Bottom line
StepAudio 2.5 TTS is best understood as a context-aware speech-performance model. Its defining features are natural-language control over delivery, inline directions for local expression, streaming synthesis, and zero-shot voice cloning from a short reference recording.
Its main trade-off is specialization. The model can provide detailed control over expressive audio, but it does not replace a general language model, and its 1,000-character input limit makes segmentation necessary for longer material. For authorized narration, character voices, emotionally aware agents, and interactive speech, those trade-offs may be worthwhile. For general reasoning, coding, tool use, or long unrestricted documents, another type of model is more suitable.