Fun-ASR: End-to-End Speech Recognition Large Language Model Fun-ASR: 端到端语音识别大模型

Chinese Dialect Recognition 中文方言识别

Supports Wu, Cantonese, Min, Hakka, Gan, Xiang, and Jin dialects, covering over 20 regions including Henan, Shaanxi, Hubei, Sichuan, Chongqing, Yunnan, Guizhou, Guangdong, Guangxi, Hebei, Tianjin, Shandong, Anhui, Nanjing, Jiangsu, Hangzhou, Gansu, Ningxia, etc.

支持的方言包括吴语、粤语、闽语、客家话、赣语、湘语、晋语，支持的口音官话覆盖中原、西南、冀鲁、江淮、兰银、胶辽、东北、北京、港台等，包括河南、陕西、湖北、四川、重庆、云南、贵州、广东、广西、河北、天津、山东、安徽、南京、江苏、杭州、甘肃、宁夏 等 20 多个地区。

Multilingual Recognition 多语言识别 (自由说)

Supports 31 languages including Japanese, Vietnamese, and code-switching scenarios.

支持的语种包括31种语言：日语、越南语等，支持语言自由切换场景。

Robust Far-Field Recognition 远场复杂背景识别

Optimized for far-field pickup and high-noise scenarios (such as conference rooms, in-vehicle environments, industrial sites), significantly improving speech recognition stability and accuracy in complex acoustic conditions.

针对远距离拾音及高噪声场景（如会议室、车载环境、工业现场等）进行深度优化，显著提升复杂声学条件下的语音识别稳定性与准确率。

Industry Customization Capability 行业定制化能力

In industrial ASR applications, personalized customization is an essential technology. Customization refers to applying additional probability preferences to specific words/phrases (such as names, places, brands, professional terms) during the recognition process, thereby significantly improving their recall rate while minimizing the impact on general recognition accuracy.

在ASR的工业落地中，个性化定制是必不可少的技术。所谓定制化，是指在识别过程中对特定词/短语（如人名、地名、品牌、专业术语等）施加额外概率偏好，从而显著提高它们的识别召回率，同时尽量不损伤通用识别准确率。

Our solution introduces RAG (Retrieval-Augmented Generation) mechanism: building a dedicated RAG library with user-configured customized words, dynamically extracting relevant vocabulary based on CTC first-pass decoding results, and precisely injecting only relevant vocabulary into the LLM's prompt to avoid interference from unrelated information. This approach expands the number of customized context words to over a thousand without increasing inference complexity, while maintaining high customized recognition performance.

我们的解决方案：引入RAG（检索增强生成）机制，将用户配置的定制词构建成专属RAG库，依据CTC第一遍解码结果动态抽取相关词汇，仅将相关词汇精准注入LLM的Prompt中避免无关信息干扰。该方案在不增加推理复杂度的前提下，将定制化上文数量扩充到上千个以上，并且保持较高的定制化识别效果。

Music & Lyrics Transcription 歌词识别

Enhanced speech recognition performance in music background interference and first-time support for accurate recognition of lyrics in songs, expanding the application potential of speech technology in entertainment and consumer scenarios such as music, short videos, and smart speakers.

强化在音乐背景干扰下的语音识别性能，并首次支持对歌曲中歌词内容的精准识别，拓展语音技术在音乐、短视频、智能音箱等娱乐与消费场景的应用潜力。

Streaming ASR 流式识别

Streaming ASR with low first-character latency, optimized theoretical first-character output delay to 160ms. While maintaining the ability to process audio streams segment by segment and return intermediate results with final transcribed text in real-time, it truly achieves millisecond-level response experience of "instant text output while speaking".

流式识别低首字延迟，首字输出理论延迟优化至160ms，在保持音频流逐段输入、实时返回中间结果与最终转写文本能力的同时，真正实现"说话即出字"的毫秒级响应体验。

Performance Benchmarks 性能基准测试

Fun-ASR achieves industry-leading performance on multiple public datasets and industrial test sets. The following are detailed performance comparison data.

Fun-ASR 在多个公开数据集和工业测试集上均达到业界领先水平，以下为详细性能对比数据。