StepAudio 2.5 ASR - An automatic speech recognition model launched by StepAudio. - AiBoss

What is StepAudio 2.5 ASR?

StepAudio 2.5 ASR is a new generation of automatic speech recognition model launched by StepAudio, designed specifically for speech transcription, meeting minutes, and long audio processing. The model adopts an architecture of Audio Encoder + Linear Adapter + 4B LLM + MTP-5, introducing multi-token prediction technology to the speech recognition field for the first time, achieving ultra-fast inference of 500 TPS. The model supports input formats such as OGG, mp3, wav, and PCM, reuses a 32K context window, and can transcrib 30 minutes of audio end-to-end in a single operation, completely eliminating the context breakage problem caused by traditional slicing solutions, achieving industry-leading speed and accuracy.

Main functions of StepAudio 2.5 ASR

High-speed speech transcriptionThe model's inference peak reaches 500 tokens/s, and 5 minutes of audio and video can be transcribed in seconds.
Long audio end-to-end recognitionReuse a 32K context window to transcribe 30 minutes of audio in a single complete transcription.
Multi-format audio supportIt is compatible with common audio formats such as OGG, mp3, wav, and PCM.
Chinese, English and dialect recognitionIt covers major Chinese and English languages, and supports dialects, Mandarin with accents, and a small number of Japanese and Arabic languages.
High-throughput decodingBased on the MTP-5 architecture, throughput is increased by 400% and inference cost is reduced by 80%.

The technical principles of StepAudio 2.5 ASR

Overall architecture designThe model adopts a four-layer stacked architecture of Audio Encoder + Linear Adapter + 4B LLM + MTP-5, which takes into account audio representation extraction, dimension alignment, language modeling and high-throughput decoding.
Audio coding layerThe 0.6B Transformer audio encoder converts the input audio into a 12.5 Hz audio embedding, completing the initial mapping from the original acoustic signal to the semantic vector.
Dimensional alignment layerThe Linear Adapter is responsible for aligning the audio encoder output with the LLM hidden dimension, enabling bridging and adaptation across modal feature spaces.
Language modeling backboneThe 4B parameter LLM serves as the core of the model, undertaking the tasks of context understanding and autoregressive text generation, and reusing the native 32K context window to achieve end-to-end modeling of long audio.
MTP Multi-Token PredictionIntroducing the same MTP-5 module as Step 3.5 Flash, a single forward propagation can additionally predict 5 candidate tokens in parallel, breaking through the speed bottleneck of traditional ASR generation by token sequentially.

How to use StepAudio 2.5 ASR

Online experienceVisit the Stepfun Experience Center at https://www.stepfun.com/studio/audio?tab=speech-recognition to upload your audio for online transcription.
API AccessVisit the Stepfun Open Platform (https://platform.stepfun.com/docs/zh/guides/models/stepaudio-2.5-asr) or the Step Plan official website (https://platform.stepfun.com/docs/zh/step-plan/integrations/audio-api) to view the API documentation and obtain the model identifier and sample code.
Demo pageVisit the Demo page https://stepaudiollm.github.io/step-audio-2.5-asr/ to view public examples and demonstrations.
Developer accessBy programmatically calling the ASR API, the recognition results can be integrated into the retrieval, summarization, quality inspection, or archiving process.

Key information and usage requirements of StepAudio 2.5 ASR

Publishing Entity: Officially released by StepFun, it is now fully available on the StepFun Open Platform and Step Plan.
Model ArchitectureIt adopts a four-layer architecture of Audio Encoder (0.6B) + Linear Adapter + 4B LLM + MTP-5, with an overall parameter scale of approximately 4.6B.
Inference performancePeak throughput reaches 500 tokens/s, throughput increases by 400%, latency decreases by 60%, and inference cost drops by 80%.
contextual abilityIt reuses the native 32K context window of LLM, and can completely transcribe up to 30 minutes of audio end-to-end in a single operation without the need for slicing and splicing.
Supported formatsCommon audio formats include OGG, mp3, wav, and PCM.
Language coverageThe game is primarily in Chinese and English, but also supports dialects, accented Mandarin, and a small amount of Japanese and Arabic.
Evaluation resultsIt achieves state-of-the-art (SOTA) performance on multiple authoritative benchmarks, including Chinese (AISHELL, Wenet, FLEURS-zh), English (LibriSpeech, Common Voice, VoxPopuli), and long audio.
Training dataPre-training uses tens of millions of hours of speech and audio data; ASR-specific training includes 100,000 hours of high-quality short audio (maximum 30 seconds) and 50,000 hours of long audio (maximum 30 minutes).
Access ChannelDevelopers can integrate through the Step Plan or the Step Plan open platform API; ordinary users can try it online at the experience center or visit the Demo Page to see the effect.

The core advantages of StepAudio 2.5 ASR

Speed BreakthroughIt was the first to introduce large language model inference acceleration technology into the field of speech recognition, with a peak inference speed of 500 tokens/s and a throughput increase of 400%.
Cost advantageInference latency is reduced by 60%, and inference cost is reduced by 80%, achieving a higher transcription limit with lower computing power.
Leading in precisionIt has achieved state-of-the-art (SOTA) level in multiple authoritative benchmarks for Chinese, English, and long audio formats, with the lowest overall error rate in the industry.
Long article stabilityReusing the 32K native context window, a single end-to-end transcription of 30 minutes of audio is performed without any loss of accuracy, completely eliminating the context breakage caused by slice splicing.
Architectural InnovationBased on the ASR+MTP-5 deep fusion architecture, it breaks through the speed bottleneck of traditional token-by-token autoregressive generation through a multi-token parallel prediction and verification mechanism.
Scene coverageSupports multiple input formats such as OGG, mp3, wav, and PCM, covering major Chinese and English languages, dialects, and complex acoustic scenarios such as Mandarin with accents.

StepAudio 2.5 ASR project address

Technical Papershttps://stepaudiollm.github.io/step-audio-2.5-asr/model-card/
Experience the demo onlinehttps://stepaudiollm.github.io/step-audio-2.5-asr/

StepAudio 2.5 ASR Competitive Comparison

Comparison Dimensions	StepAudio 2.5 ASR	Qwen3 ASR	Doubao-ASR-2603
Model Architecture	Audio Encoder+4B LLM+MTP-5	Not disclosed	Not disclosed
Reasoning speed	500 TPS, throughput increased by 400%	Standard Autoregression	Standard Autoregression
Chinese average CER	2.97%	3.17%	3.34%
English average WER	3.68%	3.85%	6.67%
Long audio WER	3.70%	4.20%	6.11%
Maximum Context	32K (30-minute end-to-end)	Unclear	Unclear
Cost optimization	Reasoning costs reduced by 80%	none	none

Application scenarios of StepAudio 2.5 ASR

Meeting minutes and interview transcriptsEnd-to-end transcription of long-duration meeting and interview recordings, maintaining terminology consistency and contextual coherence, without the need for slicing and splicing.
Subtitle generation and media asset preprocessingQuickly generate video subtitles and support batch processing of multiple audio formats such as OGG, mp3, wav, and PCM.
Long audio content archiveSuitable for complete transcription and retrieval of 30-minute audio recordings, including podcasts, lectures, and court hearings.
Real-time voice quality inspectionWith its low latency and high throughput, it is suitable for customer service call quality inspection and content compliance review scenarios.
Backend system accessThe identification results can be directly integrated into the retrieval, abstract, quality inspection, or archiving process, meeting the requirements for consistency and stability of industrial-grade terminology.