Ink 2
AI model · Speech to text · Audio / speech → Text by Cartesia
Ink 2 is a streaming speech-to-text model specifically engineered for real-time voice agents and conversational AI. It utilizes a State Space Model architecture to achieve high throughput and low latency, with a time-to-final-transcript of approximately 0.1 seconds. The model features built-in turn detection and semantic endpointing, allowing it to understand when a speaker has finished based on meaning rather than just silence. Ink 2 is designed to handle real-world audio challenges such as background noise and accented speech without requiring external filters. It is available via WebSocket
- Languages
- English, Spanish, French, Hindi, Japanese
- Input modality
- Audio (PCM, μ-law, A-law, WAV, MP3)
- Output modality
- Text (Transcripts, Turn Events)
- Time to final transcript
- 0.1s
- Supported Languages
- 5 (English, Spanish, French, Hindi, Japanese)
- Architecture
- State Space Model (SSM)
- Access Methods
- WebSocket, HTTP API, Python SDK, JS SDK
Overview
Ink 2 is a streaming speech-to-text model specifically engineered for real-time voice agents and conversational AI. It utilizes a State Space Model architecture to achieve high throughput and low latency, with a time-to-final-transcript of approximately 0.1 seconds. The model features built-in turn detection and semantic endpointing, allowing it to understand when a speaker has finished based on meaning rather than just silence. Ink 2 is designed to handle real-world audio challenges such as background noise and accented speech without requiring external filters. It is available via WebSocket for live streaming and HTTP for batch processing of recorded files.
Conversational Intelligence
- Native Turn Detection: The model features model-integrated end-of-turn detection, removing the need for external voice activity detectors.
- Semantic Endpointing: Determines the end of a turn based on the meaning of the speech rather than just silence to prevent premature interruptions.
- Turn Eager End: Provides a signal to the LLM to start generating a response before the turn is fully confirmed, reducing overall latency.
Transcription Accuracy
- Structured Data Recognition: Accurately transcribes structured data including phone numbers, emails, dates, currencies, and UUIDs.
- Noise Robustness: Maintains accuracy on noisy phone lines and in environments with background noise without requiring separate filters.
- Keyterm Prompting: Allows users to bias recognition toward specific names and industry jargon by passing up to 100 terms per connection.
Pricing
- Ink-2 usage
- 3 credits per second of audio
- Free Plan included usage
- ~1h 51m per month
- Pro Plan included usage
- ~9h 16m per month for $5/mo
- Startup Plan included usage
- ~115h 44m per month for $49/mo
Specifications
- Input modality
- Audio (PCM, μ-law, A-law, WAV, MP3)
- Output modality
- Text (Transcripts, Turn Events)
- Time to final transcript
- 0.1s
- Supported Languages
- 5 (English, Spanish, French, Hindi, Japanese)
- Architecture
- State Space Model (SSM)
- Access Methods
- WebSocket, HTTP API, Python SDK, JS SDK
Added by AntSpace from public sources (cartesia.ai).