Step TTS 2

What is Step TTS 2?

Step TTS 2 is StepFun’s dedicated text-to-speech model. Its canonical model identifier is step-tts-2. Rather than generating written answers like a general-purpose language model, it converts supplied text into spoken audio.

This positioning matters when selecting an AI model. A TTS system is used after an application has determined what should be said. For example, an assistant may generate a response with a language model and then pass that response to Step TTS 2 to produce an audio reply. Similarly, a publishing workflow can send a script to the model to create narration for a video or accessibility feature.

Step TTS 2 sits in StepFun’s catalog as a specialized audio-generation model. The supplied documentation and StepFun audio tooling identify the model line with streaming synthesis, system voices, and voice-cloning workflows.

Inputs, outputs, and supported modalities

The model’s output is direct audio, so it is appropriate for applications that need a playable or streamable voice result. The available research does not document image input, video input, speech recognition, or audio-understanding features. A recording used for voice cloning should therefore be understood as a reference asset in the cloning workflow, not evidence that the model is a general audio-understanding system.

Streaming and voice control

Streaming synthesis allows an application to receive generated speech progressively instead of waiting for the entire audio response to finish. This is useful for conversational interfaces and other experiences where reducing perceived latency is more important than producing one complete file before playback begins.

StepFun’s audio tooling also documents system voices and personalized voice workflows. For voice cloning, reference recordings are uploaded through the provider’s file API before being used in the synthesis process. The supplied research does not specify the precise audio duration, file format, language coverage, or consent requirements for cloned voices, so those details should be checked in the applicable StepFun documentation.

API access and model identifier

Applications use step-tts-2 as the model ID when accessing the model through StepFun’s developer platform. StepFun’s official audio repositories include TTS command-line and SDK tooling, and the model is included in the provider’s audio workflow.

StepFun operates China and international platform environments. Availability, quotas, regional endpoints, account requirements, and pricing may differ between those environments. The presence of a model in a tool or catalog should not be treated as a guarantee that every account has identical access. Developers should verify access in the relevant StepFun console and test the endpoint with representative text before committing to a production integration.

Pricing and usage cost

Step TTS 2 uses character-based pricing rather than the token-based pricing commonly associated with language models. Current third-party catalog information lists an approximate global rate of $0.40 per 10,000 characters. The same source lists an approximate China-region rate of ¥2.80 per 10,000 characters.

These figures should be treated as current catalog references rather than an unconditional provider guarantee. The applicable price can depend on region, account, endpoint, or changes to StepFun’s platform tariff. Confirm the rate in the StepFun console before estimating operating costs.

Character-based billing can make usage easier to estimate for scripts, announcements, and other text-controlled workloads: an application can count the characters it plans to synthesize and multiply by the applicable regional rate. However, the final cost still depends on the provider’s definition of billable characters and any account-specific terms, which are not detailed in the supplied research.

Capabilities and limitations

Step TTS 2’s strongest capability is focused speech synthesis. It is designed to produce spoken audio, supports streaming, and provides access to system voices and voice-cloning workflows. Those features make it more directly suited to voice production than a general language model that only returns text.

It should not be evaluated as a general reasoning or coding model. The supplied catalog does not document it as a model for complex reasoning, software development, image generation, video generation, embeddings, speech recognition, audio understanding, or tool calling. It also does not provide a conventional context-window or maximum-output-token specification. Those fields are either not applicable to this TTS endpoint or are not publicly documented in the supplied sources.

For the same reason, a request that requires structured text output, function calls, web search, or multi-step planning should normally be handled by another component. Step TTS 2 can potentially serve as the speech-output stage after that work is complete, but it is not documented as the component that performs it.

Reasoning, coding, and tool support

• Reasoning: Not a documented purpose of the model. It synthesizes speech from text rather than solving problems as a general reasoning model.
• Coding: Not a documented coding model. It can read text supplied for synthesis, but that does not make it a software-development assistant.
• Tool or function calling: Not documented as a model capability. Voice-cloning workflows use StepFun’s surrounding file and audio tooling, which is different from the model independently calling application tools.
• JSON or structured output: Not documented as a distinct capability. The primary result is spoken audio.

Speed, cost, and practical trade-offs

Step TTS 2’s streaming support is its clearest latency-oriented feature. Progressive audio delivery can make an assistant feel more responsive than a workflow that waits for a complete recording. The model is also specialized, so it avoids using a general-purpose model for a task that only requires speech synthesis.

Its trade-off is scope. A dedicated TTS endpoint does not replace a language model, transcription system, or multimodal model. An application that needs dialogue planning, factual reasoning, code generation, or speech recognition will need additional components or a different model type. The total system cost and latency will therefore depend on the upstream text-generation or orchestration layer as well as the TTS request.

The supplied editorial catalog rates its speed as relatively strong and its cost as moderate, but those are comparative editorial assessments, not provider-published benchmark results. No controlled latency benchmark, audio-quality score, or independent evaluation is supplied here, so those qualities should be validated with the voices, text lengths, and network conditions relevant to a particular application.

Best use cases

• Voice assistants: Convert generated assistant responses into speech, especially when progressive playback is useful.
• Narration: Produce spoken versions of scripts, articles, educational material, or other prepared text.
• Video voice-over: Create an audio track from a supplied script for videos and presentations.
• Accessibility: Add spoken output to interfaces or content where audio access is helpful.
• Conversational audio: Supply the speech layer for applications whose text responses are generated elsewhere.
• Personalized voice workflows: Use supported reference-audio processes when a suitable cloned voice is required and the relevant permissions are in place.

When to choose Step TTS 2

Choose Step TTS 2 when the central requirement is turning text into spoken audio through StepFun’s platform, particularly if streaming synthesis, system voices, or voice cloning are important. It is a sensible fit for teams that already use StepFun’s audio tooling and want a model with a clear TTS role rather than a broad, general-purpose interface.

Another option may be more appropriate when the primary task is generating the text itself, understanding an uploaded recording, transcribing speech, producing images or video, or coordinating external tools. In those cases, a language, speech-recognition, multimodal, or orchestration model should handle the relevant task, with Step TTS 2 optionally used afterward to vocalize the final text.

Bottom line

Step TTS 2 is best understood as a focused StepFun speech-generation endpoint, not as a conventional text-output LLM. Its documented strengths are text-to-speech conversion, streaming synthesis, system voices, and voice-cloning workflows. Its limitations are equally important: general reasoning, coding, multimodal understanding, tool calling, context limits, and maximum output-token limits are not documented for this model. Confirm regional pricing and access before deployment, then evaluate the generated voices and streaming behavior using the exact content and user experience the application will serve.