🔥Recentcollected in 17m

Qwen upgrades to Fun-ASR-Realtime speech recognition model

PostLinkedIn
🔥Read original on 36氪

💡New sub-100ms latency speech model from Alibaba, ideal for building high-performance real-time AI agents.

⚡ 30-Second TL;DR

What Changed

Sub-100ms latency for real-time streaming

Why It Matters

This update significantly lowers the barrier for building responsive, multilingual voice-based AI agents, improving real-time interaction quality.

What To Do Next

Integrate Fun-ASR-Realtime into your voice-agent pipeline to test its latency improvements against current STT solutions.

Who should care:Developers & AI Engineers

Key Points

  • Sub-100ms latency for real-time streaming
  • Supports 30 languages and 16 dialects
  • Accuracy performance comparable to offline models
  • Optimized for low-latency AI voice interaction

🧠 Deep Insight

AI-generated analysis for this event.

🔑 Enhanced Key Takeaways

  • The FunASR framework utilizes a Paraformer-based architecture, which is a non-autoregressive model designed specifically to balance high-speed inference with robust recognition accuracy.
  • Alibaba has integrated this model into the ModelScope open-source ecosystem, allowing developers to fine-tune the speech recognition capabilities on custom datasets.
  • The model employs a specialized VAD (Voice Activity Detection) module that operates in tandem with the streaming ASR to filter background noise and silence in real-time.
  • It supports advanced features such as timestamp prediction and hotword customization, enabling users to improve recognition accuracy for domain-specific terminology.
  • The underlying technology leverages Alibaba's DAMO Academy's research in industrial-grade speech processing, specifically focusing on reducing computational overhead for edge deployment.
📊 Competitor Analysis▸ Show
FeatureFunASR (Qwen)OpenAI WhisperDeepgram Nova-2
LatencySub-100ms (Streaming)High (Batch-optimized)Sub-300ms (Streaming)
ArchitectureParaformer (Non-autoregressive)Transformer (Autoregressive)Proprietary End-to-End
Open SourceYesYesNo (API-only)
Language Support30+ Languages/Dialects100+ Languages100+ Languages

🛠️ Technical Deep Dive

  • Architecture: Utilizes a non-autoregressive Paraformer model which eliminates the need for sequential decoding, significantly reducing latency compared to traditional RNN-T or Transformer-based models.
  • Streaming Mechanism: Implements a chunk-based processing strategy where audio is segmented into small frames, allowing the model to output partial transcripts before the full audio stream is complete.
  • Optimization: Incorporates model quantization (INT8/FP16) and ONNX Runtime support to ensure high throughput on CPU-based inference environments.
  • Training Data: Trained on tens of thousands of hours of multi-domain audio data, including noisy environments and varied acoustic conditions to ensure generalization.

🔮 Future ImplicationsAI analysis grounded in cited sources

Qwen will likely achieve dominance in Chinese-language real-time edge AI applications.
The combination of sub-100ms latency and native support for 16 dialects provides a significant competitive advantage in the domestic Chinese market over global models.
Integration of FunASR into Qwen's multimodal agents will reduce latency in voice-to-voice AI interactions.
By minimizing the ASR bottleneck, the end-to-end latency for voice-based LLM agents will be primarily limited by the LLM inference speed rather than speech processing.

Timeline

2022-09
Alibaba DAMO Academy releases the initial FunASR open-source framework.
2023-05
Introduction of the Paraformer model architecture to the FunASR ecosystem.
2024-02
FunASR reaches significant adoption milestones on ModelScope with expanded multilingual support.
2026-07
Qwen officially integrates Fun-ASR-Realtime for low-latency streaming capabilities.
📰

Weekly AI Recap

Read this week's curated digest of top AI events →

👉Related Updates

AI-curated news aggregator. All content rights belong to original publishers.
Original source: 36氪