What StepAudio 2.5 ASR is
StepAudio 2.5 ASR is StepFun’s dedicated automatic speech recognition (ASR) model. ASR systems listen to spoken audio and produce written text, making them useful for applications such as meeting transcripts, interview processing, subtitle creation, searchable voice archives, and voice-agent input.
The model’s canonical API identifier is stepaudio-2.5-asr. It occupies a specialized position in StepFun’s model catalog: it is built for transcription rather than general-purpose text generation or multimodal conversation. StepFun announced the model on April 24, 2026, and its official materials state that it is available through the StepFun Open Platform and Step Plan.
Chinese and English are the primary languages identified for the model. StepFun’s model materials also describe work intended to improve robustness across accents, dialects, noisy environments, and selected additional languages, but detailed language-by-language guarantees were not established in the supplied documentation.
Why the model is designed for speed
StepAudio 2.5 ASR combines an audio encoder, a linear adapter, a language-model backbone, and MTP-5 decoding. In simple terms, the audio encoder converts speech into information the model can process, while the language-model component turns that information into a sequence of transcript tokens.
MTP-5 refers to multi-token prediction. Instead of generating every token strictly one after another, the system can predict and verify several candidate tokens in parallel. This is intended to increase decoding throughput and reduce the delay before transcription results become available.
StepFun reports throughput of up to 500 tokens per second, approximately 400% higher throughput, 60% lower latency, and an 80% reduction in inference cost compared with its earlier ASR systems. These are provider-reported comparisons, not independent benchmark results. Actual performance can vary with hardware, audio quality, language, recording length, endpoint configuration, and account limits.
For a transcription workflow, speed matters in two different ways. Batch users may process a large archive more quickly, while interactive applications may receive partial results sooner. The model’s streaming API support is therefore relevant to applications that need incremental recognition rather than waiting for a complete recording to finish.
Long-form audio and the 32K context window
StepAudio 2.5 ASR has a reported 32K context window and is designed to transcribe up to 30 minutes of audio in one pass. This is important for recordings in which terminology, names, and conversational context continue across many minutes.
Processing a long recording in one pass can also reduce the engineering work associated with splitting audio into short segments and joining the resulting transcripts. Segmenting can introduce duplicated words, missing words at boundaries, inconsistent punctuation, or loss of context. The model’s long-form design does not eliminate every transcription challenge, but it is intended to reduce the need for such manual stitching.
The 30-minute figure should be treated as the documented design target rather than an unconditional guarantee for every request. File size, account restrictions, API settings, audio format, and service-side limits may affect what can be submitted successfully. The supplied research does not establish a separate maximum output-token limit.
Supported inputs and outputs
| Capability | StepAudio 2.5 ASR |
|---|---|
| Primary input | Audio |
| Primary output | Transcription text |
| Audio formats identified in official materials | PCM, OGG, MP3, and WAV |
| Context window | 32K |
| Designed maximum audio length | Up to 30 minutes in one pass |
| Streaming | Available through server-sent event ASR usage |
| Speech generation | Not supported |
| Image or video input | No official support identified |
The model returns text transcription, not audio. It does not natively synthesize speech, generate music, create images, or generate video. It should therefore be treated as the recognition component of a voice pipeline, not as a complete speech-to-speech assistant.
The available research does not verify built-in diarization, detailed timestamp output, structured JSON output, or a fine-tuning interface. If an application depends on speaker labels, word-level timing, schema-constrained responses, or custom training, those capabilities should be confirmed separately before implementation rather than assumed from the model’s general ASR function.
Pricing and service access
StepFun pricing references list StepAudio 2.5 ASR at ¥0.15 per audio hour through the Step Plan channel. This is a duration-based price: it is associated with the amount of audio processed, not the number of input or output text tokens.
The listed price provides a useful indication of the model’s cost positioning for transcription workloads, especially when compared with services priced on token volume or more expensive real-time speech infrastructure. However, availability, quotas, regional endpoints, billing rules, and commercial terms can change. Teams should verify the current price and eligibility in the StepFun account and pricing documentation before budgeting a production deployment.
The official API documentation identifies streaming server-sent event usage for ASR requests. That makes the service relevant to applications that display a live or progressively updated transcript, although the supplied research does not specify every streaming parameter or guarantee a particular time-to-first-result.
Reasoning, coding, and tool support
StepAudio 2.5 ASR is not a reasoning model in the usual sense. Its job is to recognize speech and produce a transcript, not to solve multi-step problems, interpret instructions, or independently analyze a conversation. Any summarization, question answering, extraction, or workflow decision should normally be handled by a separate text model after transcription.
Similarly, the model has no established coding capability and is not intended to generate or review software. It does not provide general-purpose tool or function calling. The model can serve as an upstream input component for a larger application—for example, converting a spoken request into text before another system interprets it—but that downstream orchestration is separate from StepAudio 2.5 ASR itself.
These distinctions matter when comparing an ASR model with a general multimodal assistant. A broader model may accept audio and reason about its contents, but a specialized recognizer can be a better fit when the required result is a transcript and the priorities are processing speed, predictable audio handling, and cost per recording hour.
Main strengths and limitations
Strengths
- Fast reported decoding: StepFun claims up to 500 tokens per second, with substantial throughput and latency improvements over earlier StepFun ASR systems.
- Long-form support: The 32K context window and up-to-30-minute single-pass design are useful for meetings, interviews, lectures, podcasts, and similar recordings.
- Production-oriented access: The model has a canonical API identifier, Open Platform availability, Step Plan access, and documented streaming usage.
- Audio-duration pricing: The published ¥0.15 per audio hour reference is straightforward for estimating transcription costs based on recording length.
- Focus on Chinese and English: The model is specifically positioned for these languages, with additional robustness claims covering accents, dialects, and noisy conditions.
Limitations
- Not a general assistant: It produces transcription text and should not be selected for open-ended chat, reasoning, coding, or general multimodal interaction.
- No native speech output: It cannot turn text back into speech or conduct speech-to-speech conversations by itself.
- Unverified advanced transcript features: Official evidence supplied here does not establish diarization, detailed timestamps, JSON schema output, or fine-tuning.
- Provider-reported performance: The speed, latency, and cost improvements are StepFun claims and may not match every deployment.
- Operational details remain account-dependent: Quotas, regional access, endpoint behavior, and current pricing should be checked before production use.
Best use cases
StepAudio 2.5 ASR is a strong candidate when the central task is converting substantial amounts of speech into text quickly. Suitable examples include:
- Transcribing Chinese- or English-language meetings and interviews
- Creating first-pass subtitles for recorded video
- Processing lectures, podcasts, and other long recordings
- Building searchable archives of calls or voice recordings
- Adding speech recognition to voice-agent pipelines
- Supporting near-real-time transcription through streaming results
It is especially attractive when a workflow values throughput and audio-duration cost more than broad reasoning capabilities. For example, a media team processing many hours of interviews may benefit more from a specialized ASR service than from sending every recording to a general multimodal model.
When to choose this model
Choose StepAudio 2.5 ASR when you need a focused transcription engine, support for relatively long recordings, Chinese or English recognition, and an API that can provide streaming results. Its reported speed and duration-based pricing make it most compelling for high-volume or latency-sensitive speech-to-text workloads.
Another option may be more appropriate when the application must understand images or video directly, generate spoken responses, perform complex reasoning on audio, identify speakers with verified diarization, return strictly structured output, or combine transcription and conversation in one model. In those cases, a broader multimodal or speech-to-speech system may reduce integration work even if it has different cost or performance trade-offs.
A practical architecture can also combine StepAudio 2.5 ASR with another model: use StepAudio for recognition, then pass the transcript to a separate language model for summarization, extraction, search, or agent actions. This preserves the specialized model’s transcription role while avoiding assumptions that it can perform downstream reasoning or tool execution.
Bottom line
StepAudio 2.5 ASR is a specialized StepFun model for fast, long-form speech recognition. Its clearest advantages are the reported MTP-5 decoding efficiency, 32K context, up-to-30-minute single-pass transcription design, streaming API access, and audio-hour pricing reference. Its boundaries are equally important: it outputs text, does not generate speech, and is not documented as a general reasoning, coding, diarization, or structured-output model. For applications whose primary requirement is efficient Chinese and English transcription, it is a focused option; for broader audio understanding or conversational behavior, it should be paired with—or replaced by—a model designed for those tasks.