nvidia / canary-qwen-2.5b

🎤The best open-source speech-to-text model as of Jul 2025, transcribing audio with record 5.63% WER and enabling AI tasks like summarization directly from speech✨

  • Public
  • 41 runs
  • GitHub
  • Weights
  • Paper
  • License
Iterate in playground

Run time and cost

This model runs on Nvidia L40S GPU hardware. We don't yet have enough runs of this model to provide performance information.

Readme

Canary-Qwen-2.5B: Speech Recognition + AI Analysis 🎙️🧠

You know how most speech AI either just writes down what you said or completely misses the context? Canary-Qwen-2.5B actually gets both. Upload a 2-hour meeting recording and get perfect transcription with timestamps. Then ask it “What were the main decisions made?” and it analyzes the content like a human would—understanding context, extracting insights, and reasoning through complex discussions step by step.

Note:
The model weights are for commercial and non-commercial use under NVIDIA’s CC-BY-4.0 license.
Attribution is required when using the model.

What Canary-Qwen-2.5B does ✨

Canary-Qwen-2.5B transforms how you analyze audio by: - Perfect transcription: Goes beyond basic speech-to-text to understand punctuation, capitalization, and conversation flow - Smart reasoning: Shows its thinking process when analyzing complex discussions or presentations
- Dual capability: Handles pure transcription and intelligent analysis equally well - Long-form processing: Analyzes up to 2 hours of audio content in context - Question answering: Answer any question about your audio content - Professional quality: Provides detailed analysis suitable for business and research

Audio analysis capabilities 🎵

Canary-Qwen-2.5B uses advanced reasoning to understand not just what words are spoken, but their relationships, context, and meaning. It can analyze everything from casual conversations to formal presentations.

Key features:

🧠 Contextual understanding for detailed conversation analysis
🎙️ Speaker awareness with context and discussion flow recognition
🎼 Content structure including topics, decisions, and action items
🔊 Audio quality handling for real-world recordings
📝 Flexible questioning - ask it anything about your audio
⏱️ Extended processing up to 2 hours of continuous audio
🎯 Timestamp precision for focusing on specific segments

How to get the best results 🌟

Basic approach: - Canary-Qwen works with MP3, WAV, M4A, FLAC, OGG, AAC, anything - Ask clear questions about what you want to know
- Get detailed insights and analysis

Advanced control: - Use specific prompts to customize the analysis format - Enable reasoning mode for step-by-step thinking - Include timestamps for detailed segment analysis - Use targeted questions for focused insights

Example questions:

For business meetings: - “What were the main action items and who is responsible for each?” - “Summarize the key decisions made and any concerns raised” - “Who were the speakers and what were their main points?”

For interviews and podcasts: - “What are the most interesting insights shared by the guest?” - “Create bullet points of the main topics discussed”
- “What questions did the host ask and how did the guest respond?”

For lectures and presentations: - “Break down the main concepts explained in this lecture” - “What examples were used to illustrate key points?” - “Create a structured summary with headings and key takeaways”

Parameter controls 🎛️

include_timestamps (true/false): Adds time markers to transcription. When enabled, shows exactly when each segment was spoken.

show_confidence (true/false): Reveals the model’s reasoning process. Shows how it arrived at its analysis and conclusions.

llm_prompt: Your question or instruction for analysis. Be specific about what insights you want.

What makes Canary-Qwen-2.5B special 🚀

Traditional speech models are limited to basic transcription or simple commands. Canary-Qwen-2.5B changes this by:

  • True dual capability: Combines NVIDIA’s state-of-the-art speech recognition with Qwen’s language understanding in one seamless model
  • Advanced reasoning: Doesn’t just transcribe—analyzes content, understands context, and answers complex questions
  • Unified approach: No need to use separate tools for transcription and analysis
  • Professional interface: Works seamlessly through Replicate’s web interface
  • Enterprise quality: 5.63% word error rate puts it among the best speech recognition systems available

Best use cases 🎯

Canary-Qwen-2.5B excels at: - Meeting analysis: Extract action items, decisions, and key discussions from recordings - Content creation: Transcribe interviews, podcasts, and videos for editing and repurposing - Research processing: Analyze recorded interviews, focus groups, and qualitative data
- Educational support: Convert lectures and presentations into structured notes and summaries - Business intelligence: Process customer calls and feedback for insights and trends - Accessibility services: Create accurate transcripts with intelligent summaries for hearing-impaired users

Limitations to consider ⚠️

  • Works best with clear English speech
  • Background noise may affect transcription accuracy
  • Complex audio scenarios benefit from specific questioning
  • Processing time increases with longer audio files
  • Analysis quality depends on audio clarity and content complexity

Research background 📚

Canary-Qwen-2.5B represents a significant advance in speech-language models, built on NVIDIA’s research in speech recognition and Alibaba’s work in language understanding. It combines FastConformer speech encoding with Transformer-based language modeling.

The model builds on the Canary series, introducing: - Hybrid speech-language architecture - 2.5 billion parameter optimization - Multi-modal reasoning capabilities
- Extended context processing

Original research: Less is More: Accurate Speech Recognition & Translation without Web-Scale Data

Important licensing note 📝

This model is for commercial and non-commercial use under NVIDIA’s CC-BY-4.0 license.
Attribution is required when using the model.

The model is built on NVIDIA Canary-1B and Qwen3-1.7B which have their own open-source license requirements.

Built by NVIDIA. Packaged for Replicate by @zsxkib.

Disclaimer ‼️

I am not liable for any direct, indirect, consequential, incidental, or special damages arising out of or in any way connected with the use/misuse or inability to use this software.


⭐ Star the repo on GitHub!
🐦 Follow @zsxkib on X
💻 Check out more projects @zsxkib on GitHub