Models & versions
Every Krisp model in one place: model IDs, the SDK or API that serves them, and where to read more.
This page lists Krisp's models by product. Each row links to the page that explains how to choose and use the model.
Krisp API
You choose the model in the client hello, with the model field.
| Model ID | Endpoint | Use it for | Status |
|---|---|---|---|
krisp-realtime-stt-agents-v1 | /stt | Speech-to-text for voice AI agents: fast partials and end-of-turn events. See STT for Agents. | Beta |
krisp-realtime-stt-humans-v1 | /stt | Speech-to-text for human conversations: diarization, PII redaction, custom vocabulary. See STT for Humans. | Beta |
krisp-tts-v1 | /tts | Text-to-speech. See Text-to-Speech. | Beta |
VIVA-tel | /voice-isolation | Voice Isolation for telephony audio, 8 or 16 kHz. See Voice Isolation. | Current |
| (none) | /vt | Voice Translation has no model field. You choose the languages and the voice. See Voice Translation. | Current |
VIVA SDK
Download the models from the Developers Portal and load them as .kef files.
| Model | Feature | Use it for |
|---|---|---|
krisp-viva-vi-tel-v2.5.1 | Voice Isolation | Telephony, 8–16 kHz. Best noise cancellation. |
krisp-viva-vi-tel-lite-v2.5 | Voice Isolation | Telephony, 8–16 kHz. Lowest CPU usage. |
krisp-viva-vi-tel-v2.7 | Voice Isolation | Telephony, 8–16 kHz. The most ASR-friendly model. |
krisp-viva-vi-pro-v1 | Voice Isolation | WebRTC, above 16 kHz. |
krisp-viva-nc-ss-v1 | Noise cancellation | Smart-speaker and far-field audio. Doesn't remove background voices. |
krisp-viva-tp-v3 | Turn-Taking (TTv3) | End-of-turn prediction, all languages. Replaces krisp-viva-tt-v2. |
krisp-viva-ip-v1.1 | Interrupt Prediction | Telling interruptions from backchannels. Replaces krisp-viva-ip-v1; the files are not interchangeable between SDK versions. |
krisp-viva-vad-v2 | VAD | Voice activity detection. |
See Voice Isolation, Turn-Taking and Interrupt Prediction.
RTC SDK
The RTC SDK models are named krisp-<technology>-<stream>-<size>-v<version>. In the name, nc means noise cancellation and bvc background voice cancellation; o is the outbound (microphone) stream and i the inbound (speaker) stream. For sizes, quality, CPU and recommendations, see the Model Selection Guide.
| Model | Stream | Use it for |
|---|---|---|
krisp-nc-o-pro-v1 | Outbound | Noise cancellation, 32 kHz, high quality |
krisp-nc-o-med-v7 | Outbound | Noise cancellation, 32 kHz, for mobile and power-limited systems |
krisp-nc-o-lite-v1 | Outbound | Noise cancellation, 16 kHz, lowest CPU |
krisp-nc-o-nb-v2 | Outbound | Noise cancellation, up to 8 kHz (narrow-band) |
krisp-bvc-o-pro-v3 | Outbound | Background voice cancellation |
krisp-bvc-o-lite-v2 | Outbound | Background voice cancellation, lower CPU |
krisp-nc-i-wb-pro-v3 | Inbound | Noise cancellation, wide-band |
krisp-nc-i-nb-pro-v1 | Inbound | Noise cancellation, narrow-band |
krisp-nc-i-nb-lite-v2 | Inbound | Noise cancellation, narrow-band, lower CPU |
Large models, such as krisp-nc-o-max-v1 and krisp-bvc-o-max-v3, are available for evaluation on request. Ask your Krisp application engineer.
Accent Conversion models come with the Accent Conversion SDK packages. See Accent Conversion.
Updated 3 days ago