Fun-ASR1.5 is a large end-to-end speech recognition model launched by Alibaba Tongyi team. Based on MoE architecture, a single model supports 30 languages, seven major dialect systems and more than 20 local accents, and has built-in intelligent punctuation prediction and text normalization capabilities.

Fun-ASR1.5: Alibaba Tongyi end-to-end speech recognition model

Core parameters and statistics of Fun-ASR1.5

Project	Details
Product name	Fun-ASR1.5 (Alibaba Tongyi speech recognition model)
Product type	End-to-end speech recognition large model
Delivery Form	Python Library/API/CLI/MCP Server
Supported languages	30 languages (Chinese, English, Japanese, Korean, French, German, Spanish, Portuguese, Russian, Arabic, etc.)
Dialect coverage	Seven major dialect systems, focusing on optimizing 15 high-demand dialects such as Shanghainese, Cantonese, and Sichuan dialects
Architecture	MoE (Mixed Expert Architecture)
Latest version	FunASR 1.3.19 (2026-07-19)
Open source agreement	MIT (Toolkit) / Model weights come with separate agreement
GitHub Stars	19.3k
Target users	Developers, enterprises, research institutions

Parameter Interpretation: A single model covering 30 languages means that there is no need to maintain separate model instances for each language during deployment. The MoE architecture ensures that the corresponding expert module is only activated when a specific language is recognized, and the reasoning efficiency is higher than that of a dense model of the same scale. The dialect error rate (CER) has dropped by 56.2% compared to the previous version, which can significantly reduce manual proofreading costs in practical scenarios.

Users and market recognition of Fun-ASR1.5

Behind Fun-ASR1.5 is the open source FunASR project (GitHub 19.3k Stars) of Alibaba Damo Academy. It is currently one of the most active end-to-end toolkits in the Chinese speech recognition community. The cumulative downloads of PyPI have covered a large number of individual developers and corporate users, with more than 178 community contributors.

In the benchmark test, Fun-ASR-Nano (the flagship model of the Fun-ASR1.5 series) achieved a word error rate (CER) of only 8.20% on the Chinese long audio test set, which is much lower than Whisper-large-v3-turbo's 21.71%. The inference speed reached 340 times real-time (with vLLM), which is more than 26 times that of Whisper. The SenseVoiceSmall model can achieve 17x real-time inference on the CPU and can meet production-level throughput without the need for a GPU.

This project has been widely used in scenarios such as meeting minutes, intelligent customer service, voice search, subtitle generation, and online education. Enterprise-level users can directly call the API through the Alibaba Cloud Bailian platform.

Cost Advantages of Fun-ASR1.5

Fun-ASR1.5 is a completely open source project, and its core advantage is zero license cost.

Cost Dimension	Description
Toolkit usage	MIT license, completely free and available for commercial use
Model weight	Subject to the model card agreement, most of which are free for commercial use
API call (Alibaba Cloud Bailian)	Pay-as-you-go billing, please refer to Alibaba Cloud's real-time quotation
Private deployment	Build your own server and only bear the infrastructure cost
Competitive product comparison (Whisper API)	The inference speed is 26 times higher under self-deployment, and the hardware cost is significantly reduced under the same throughput

C-side cost: The online experience through Moda Community is completely free, and local deployment only requires a machine with a GPU (the CPU can also run the SenseVoiceSmall model). Business-end cost: Enterprises can choose to pay on a pay-as-you-go basis for Alibaba Cloud Bailian API, or deploy it privately to avoid data going out of the domain. In self-deployment mode, the inference efficiency of Fun-ASR1.5 enables it to carry much higher concurrency than Whisper on the same hardware.

Main functions of Fun-ASR1.5

• Multi-language recognition: A single model covers 30 languages including Chinese, English, Japanese, Korean, French, German, Spanish, Portuguese, Russian, and Arabic. There is no need to deploy independent models for each language. Suitable for transnational conferences and multi-language content production scenarios.

• Automatic language switching (Code-Switching): No need to preset language tags, automatically recognize and process multi-language mixed voices in the same conversation. This is extremely practical for scenarios such as business meetings and technology podcasts where Chinese and English are mixed, eliminating the tedious manual segmentation annotation.

• Dialect and Accent Recognition: Covers seven major dialect systems and more than 20 local accents, focusing on optimizing 15 high-demand dialects such as Shanghainese, Cantonese, Sichuan, and Hokkien. The error rate of dialect characters has dropped by 56.2% compared with the previous version, supporting the restoration of original dialect characters.

• Ancient Poetry Recitation Recognition: Constructed a real-person recitation corpus covering classic texts such as "The Book of Songs", "Chu Ci", Li Bai and Du Fu's Poems, Su Shi and Xin Qiji's Ci Poems, with a character-level accuracy of 97%. It has unique value to the digital scenario of Chinese studies education and cultural inheritance.

• Intelligent post-processing: Automatically insert punctuation marks such as commas, periods, question marks, etc. based on contextual semantics, and automatically convert spoken numbers, dates, amounts, phone calls, etc. into standardized written formats. This feature can significantly reduce post-production manual editing time for scenarios such as meeting minutes and legal transcripts.

Fun-ASR1.5 model and version evolution

Fun-ASR1.5 is an open source project that continues to iterate. The latest main version is FunASR 1.3.19 (released on 2026-07-19). The core evolution process is as follows:

Version	Date	Key Changes
FunASR 1.3.19	2026-07-19	Real-time WebSocket long session diagnostic logs packaged into PyPI documentation
FunASR 1.3.18	2026-07-19	Fix CLI subtitle segmentation regression, SRT output is segmented by sentences
FunASR 1.3.17	2026-07-19	SenseVoice llama.cpp runtime improvements, Windows CUDA support
FunASR 1.3.16	2026-07-18	Fun-ASR-Nano real-time WebSocket service installed directly from PyPI
FunASR 1.3.15	2026-07-17	Streaming reliability improvement, Fun-ASR-Nano compatibility fix
FunASR 1.3.3	2026-05-24	funasr-server CLI, OpenAI compatible API, MCP Server
Fun-ASR-Nano-2512	2025-12-15	The first Fun-ASR-Nano flagship model released

In terms of model pedigree, the FunASR toolkit provides multiple model choices: Fun-ASR-Nano (Chinese/English/Japanese + dialect), Fun-ASR-MLT-Nano (31 languages), SenseVoiceSmall (5 languages ​​+ emotion recognition), Paraformer (low-latency streaming ASR), Qwen3-ASR (52 languages), etc. Users can flexibly select based on scene accuracy, language coverage and hardware conditions.

Technical advantages of Fun-ASR1.5

The core technical advantages of Fun-ASR1.5 can be broken down from three levels: architecture, inference efficiency, and deployment flexibility:

• MoE Hybrid Expert Architecture: The model contains multiple language-related expert sub-networks. When a specific language is heard, only the corresponding expert module is activated. This design makes the parameter efficiency of a single model covering 30 languages ​​much higher than that of a dense model of the same scale. The amount of calculation during inference is greatly reduced. At the same time, each expert module can be optimized independently without interfering with each other.

• Graded and staged training strategy: Use accurate data for training in stages and stages - first do general multi-language pre-training, then dialect/accent fine-tuning, and finally do special scene (such as ancient poetry) alignment. This training process makes the model significantly more robust in real-world complex speech scenarios than a model trained in a single stage.

• Full-link deployment matrix: From Python library (pip install funasr) to OpenAI compatible API service (funasr-server), from llama.cpp/GGUF edge-side binary to Docker containerized deployment, from MCP Server to vLLM batch inference engine, covering the complete deployment needs from personal development machines to large-scale production clusters. In particular, the llama.cpp runtime brings FunASR to pure CPU and edge devices without the need for a Python runtime, giving it unique advantages in IoT and embedded scenarios.

• Leading in inference performance: Fun-ASR-Nano + vLLM reaches 340x real-time (RTFx), SenseVoiceSmall reaches 17x real-time on CPU. Compared with Whisper-large-v3's 13 times real-time (GPU), FunASR's inference speed on the CPU even exceeds Whisper's performance on the GPU, making CPU cardless deployment a feasible solution.

Fun-ASR1.5 How to use

Fun-ASR1.5 provides a variety of usage methods, and developers can flexibly choose according to the scenario:

Entrance	Description
Python library (pip)	pip install funasr, called via AutoModel API
CLI command line	funasr audio.wav is transliteration and supports SRT/JSON output
OpenAI compatible API	funasr-server --device cuda starts the service, POST /v1/audio/transcriptions
MCP Server	For AI Agents such as Claude/Cursor to call speech recognition capabilities
Magic Scope Community Online Experience	Use the browser directly, no installation required: https://modelscope.cn/studios/iic/FunAudio-ASR
llama.cpp binary	Pure CPU/edge run, no Python context required

Get started quickly with Python:

from funasr import AutoModel

model = AutoModel(model="FunAudioLLM/Fun-ASR-Nano-2512", device="cuda")
result = model.generate(input="audio.wav")
print(result[0]["text"])

CLI one-line transcription:

funasr audio.wav --output-format json --spk --timestamps

Start API service:

funasr-server --device cuda
# After the service is started, provide an OpenAI compatible interface at localhost:8000

Typical usage process: Install FunASR → Load model (automatically download weights) → Input audio files/streams → Obtain structured recognition results → Post-processing (punctuation/normalization). There is no need to manually set language tags in the whole process, the model automatically recognizes it.

Product Pricing for Fun-ASR1.5

Fun-ASR1.5’s open source strategy gives it significant cost advantages:

• Community Edition (Free): Install GitHub repository download through PyPI, and experience the Magic Community online, all for free. The MIT open source license allows commercial use and secondary development.
• Alibaba Cloud Bailian API (pay-as-you-go): For users who do not want to build their own infrastructure, they can call the API through the Alibaba Cloud Bailian platform and be billed based on the amount of calls. The specific price is subject to the Alibaba Cloud real-time pricing page.
• Enterprise privatized deployment (self-built): Enterprises can deploy on their own servers and only need to bear the cost of computing resources. Taking advantage of FunASR's high inference efficiency, a single GPU can handle a large number of concurrent requests.

Compared with commercial speech recognition services (such as Azure Speech, Google Cloud Speech-to-Text, Whisper API), the TCO (total cost of ownership) of the Fun-ASR1.5 self-deployment solution has significant advantages in medium and large-scale use scenarios, and the data does not go out of the domain, meeting compliance requirements.

Application scenarios of Fun-ASR1.5

• Multi-language conferences and multi-language transcription: Accurately transcribe multi-language mixed dialogue content in real-time during multinational meetings. There is no need to preset languages in advance. A single model can cover 30 languages. Automatic punctuation and text normalization significantly reduce post-processing costs.

• Intelligent Customer Service Speech Analysis: Batch transcribe customer service call recordings into structured text, support dialect recognition and speaker separation, and be used for service quality monitoring, hot issue mining and sentiment analysis.

• Online education and digitization of Chinese studies: The 97% character accuracy of ancient poetry recitation recognition makes it a unique tool for online Chinese studies education, which can translate students' recitations into text in real time and conduct comparison and scoring.

• Content production and subtitle generation: Video creators can transcribe audio into SRT subtitle files through a CLI one-line command. It supports speaker tags and timestamps, and is suitable for content production such as podcasts, course videos, and news interviews.

• Voice interaction between IoT and edge devices: When running through llama.cpp/GGUF, Fun-ASR1.5 can run on pure CPU devices and is suitable for edge scenarios such as smart speakers, car voice, and industrial control.

Applicable groups of Fun-ASR1.5

• AI application developers: Developers who need to integrate speech recognition into their own applications can quickly access it through Python SDK, REST API or MCP Server. pip install funasr can start.

• Enterprise IT and Data Team: For enterprises that need to privatize the deployment of speech recognition services, FunASR provides a complete deployment matrix (Docker, Kubernetes, OpenAI compatible API) to support data not leaving the domain.

• Research institutions and academic users: Open source model weights and a complete training and fine-tuning tool chain, suitable for research work in speech recognition, dialectology, multilingual ASR and other directions.

• Content Creators and Media Personnel: Creators who need to quickly transcribe audio/video into text and generate subtitles. The CLI can complete the transcoding with one line of commands and supports output in formats such as SRT/JSON/VTT.

• Not suitable for the crowd: For real-time communication scenarios with strict requirements for ultra-low latency (<100ms end-to-end), it is recommended to evaluate after testing; users who require pure cloud-free operation and maintenance hosting services are recommended to give priority to Alibaba Cloud Bailian API instead of self-deployment; pure business users who are not familiar with the command line or Python programming need to use Alibaba Cloud Bailian visual interface or third-party integration.

Summary and Outlook of Fun-ASR1.5

Fun-ASR1.5 is an important open source achievement of Alibaba Tongyi team in the field of speech recognition. Its core competitiveness lies in: single-model multi-language coverage (30 languages + seven major dialects) reduces the operation and maintenance complexity of multi-model deployment; the inference efficiency brought by the MoE architecture enables it to achieve throughput far exceeding that of competing products on the same hardware; the full-link open source strategy (MIT protocol + complete deployment tool chain) provides enterprises and developers with a low-cost, highly controllable speech recognition infrastructure.

Limitations: Although model weights can be used for free, the license agreements of different models are different, and the authorization terms of the specific model card need to be confirmed before commercial use; ultra-low latency streaming scenarios (such as real-time intercom translation) still require targeted tuning; for extremely low-resource edge devices (MCU level), the llama.cpp runtime is still being continuously optimized.

Procurement/Adoption Risk Assessment: As an open source project maintained by Alibaba Damo Academy, FunASR has long-term and stable community and enterprise support, with 19.3k GitHub Stars and 178 contributors proving its ecological activity. Enterprises need to pay attention when adopting: the open source version does not provide SLA guarantee, and key production enterprises are recommended to obtain enterprise-level support through Alibaba Cloud Bailian API; independent license agreements for model weights need to be confirmed one by one before commercial use; the project iteration rhythm is fast (multiple small versions are released every month), and production enterprises should lock the version and do regression testing. Overall, Fun-ASR1.5 is currently one of the most cost-effective open source options in the field of Chinese speech recognition for teams with self-building capabilities and data compliance requirements.

Related tools: , udio

Version Info

• FunASR 1.3.19 ：Added real-time WebSocket long session diagnosis function, improved CLI subtitle generation and punctuation loading, and updated community documentation. （2026-07-19）
• FunASR 1.3.18 ：Fixed SRT subtitle segmentation regression problem, supported sentence-level timestamp request, and improved punctuation model loading. （2026-07-19）
• FunASR 1.3.17 ：Added real-time WebSocket session diagnostic logs, SenseVoice llama.cpp runtime improvements, Windows CUDA support. （2026-07-19）
• FunASR 1.3.16 ：Fun-ASR-Nano real-time WebSocket service can be installed directly from PyPI and supports client-driven endpoint mode. （2026-07-18）
• FunASR v1.3.15 ：Improve streaming reliability, fix Fun-ASR-Nano compatibility issues, and improve CLI and text processing. （2026-07-17）