## Introduction FireRedASR is an industrial-grade automatic speech recognition system developed by the FireRed Team, delivering state-of-the-art performance on public Mandarin ASR benchmarks while supporting English and Chinese dialect recognition. With 1,800+ GitHub stars and an Apache-2.0 license, FireRedASR distinguishes itself by offering two complementary model architectures that balance cutting-edge accuracy against practical deployment efficiency. Automatic speech recognition has become a critical component in voice assistants, transcription services, and accessibility tools. While OpenAI's Whisper popularized open-source ASR, FireRedASR pushes the accuracy frontier significantly further, achieving a 3.05% average Character Error Rate (CER) across major Mandarin benchmarks, substantially outperforming previous open-source alternatives. ## Architecture and Models FireRedASR ships two model variants designed for different deployment scenarios: | Model | Parameters | Architecture | Focus | |-------|-----------|-------------|-------| | FireRedASR-LLM | 8.3B | Encoder-Adapter-LLM | Maximum accuracy (SOTA) | | FireRedASR-AED | 1.1B | Attention Encoder-Decoder | Efficiency-accuracy balance | **FireRedASR-LLM** (8.3B parameters) uses an Encoder-Adapter-LLM framework that combines a speech encoder with a large language model through an adapter layer. This architecture leverages the LLM's language understanding to achieve state-of-the-art recognition accuracy, particularly excelling in noisy environments, accented speech, and domain-specific vocabulary. **FireRedASR-AED** (1.1B parameters) employs a traditional Attention-based Encoder-Decoder architecture optimized for production deployment. At roughly one-eighth the size of the LLM variant, it delivers competitive accuracy while requiring significantly less compute and memory. ## Key Capabilities **State-of-the-Art Mandarin ASR**: Achieves an average CER of 3.05% across four major benchmarks (AISHELL-1, AISHELL-2, WenetSpeech-net, WenetSpeech-meeting), setting new records on each. **FireRedASR2S All-in-One System**: The latest February 2026 release integrates Voice Activity Detection (VAD), Language Identification (LID), punctuation restoration, and ASR into a single unified pipeline, achieving SOTA across all components. **Singing Lyrics Recognition**: A unique capability not found in most ASR systems: FireRedASR can accurately transcribe singing lyrics, making it valuable for music technology and karaoke applications. **Multi-Language Support**: Handles Mandarin, multiple Chinese dialects (Cantonese, Shanghainese, Sichuan dialect, etc.), and English with a single model deployment. **Batch Processing**: Built-in support for batch inference with command-line tools and a Python API, enabling efficient processing of large audio datasets. **Production-Ready Design**: Industrial-grade reliability with comprehensive error handling, logging, and monitoring support designed for enterprise deployment scenarios. ## Limitations FireRedASR's primary optimization target is Mandarin and Chinese dialects, meaning its English recognition, while functional, does not match dedicated English ASR models like Whisper Large V3. The LLM variant at 8.3B parameters requires significant GPU memory, making the 1.1B AED model more practical for most deployments. The project documentation and community discussions are predominantly in Chinese, which may limit accessibility for international developers. The singing lyrics recognition feature, while innovative, works best with Chinese-language music. Real-time streaming inference is not yet a primary focus, with the system optimized more for batch and near-real-time processing. ## Who Should Use This FireRedASR is the top choice for developers building Mandarin-focused speech recognition applications requiring the highest possible accuracy. Enterprise teams processing Chinese audio content at scale benefit from the AED model's efficiency-accuracy balance. Music technology companies exploring lyrics transcription will find the singing recognition capability uniquely valuable. Teams building multilingual ASR pipelines covering both Chinese and English can use FireRedASR as their primary Chinese ASR component alongside dedicated English models. Researchers studying encoder-LLM fusion architectures for speech gain access to a well-documented, industrial-grade implementation.