Cloud Transcription Providers

Cloud transcription securely streams real-time audio from your Mac to the configured third-party speech API. Before connecting, create an account with the provider, review their pricing and data policies, and generate a dedicated API key for Kapinote.

Connecting a provider

1. Open Settings → Speech and choose a cloud speech model.
2. Click Configure and paste your API key.
3. Keep the default Model and Base URL unless your organization uses a private endpoint or dedicated proxy.
4. Configure language hints, speaker diarization, and endpointing options as needed.
5. Save and run the built-in test to verify network and authentication connectivity.

API Key Tip: Ensure there are no leading or trailing whitespace characters or quotation marks around your key. If verification fails, verify that your key has real-time speech permissions and remaining account balance.

1. Soniox

• API Key: Follow the Soniox getting-started guide.
• Defaults: Model is set to stt-rt-v4 and Base URL to wss://stt-rt.soniox.com/transcribe-websocket.
• Language Hints: Supports one or multiple language hints to guide recognition without strictly restricting non-hinted languages. See Soniox language hints documentation.
• Features: Supports high-precision speaker diarization and endpoint detection.
• Languages: Kapinote lists the most common languages; Soniox officially supports 60+ languages. See the Soniox supported languages list.

2. Deepgram

• API Key: Generate a Project Key in the Deepgram Console. See Deepgram API Key Guide.
• Defaults: Default model is Nova-3 at wss://api.deepgram.com/v1/listen.
• Languages: Supports explicit single-language selection or automatic multilingual detection.
• Keyterm Prompting: Kapinote passes up to 100 custom vocabulary terms (up to 50 characters each). Nova-3 leverages Deepgram's Keyterm Prompting to boost specialized technical terms.

3. AssemblyAI

• API Key: Copy your API key from the AssemblyAI dashboard.
• Defaults: Default model is universal-streaming-multilingual at wss://streaming.assemblyai.com/v3/ws.
• Languages: Supports automatic detection or explicit selection of English, Spanish, French, German, Italian, or Portuguese.
• Vocabulary: Passes up to 100 custom terms (up to 50 characters each).
• Consult the AssemblyAI Streaming API reference for technical specifications.

4. ElevenLabs

• API Key: Create an API key in the ElevenLabs console.
• Defaults: Uses scribe_v2_realtime at wss://api.elevenlabs.io/v1/speech-to-text/realtime.
• Languages: Supports automatic detection or specific target language selection.
• Vocabulary: Passes up to 50 custom terms (up to 20 characters each).
• See the ElevenLabs Speech to Text overview and Real-time API reference.

Language configuration differences

Speech providers handle multilingual recognition differently: Soniox takes multiple hints simultaneously, Deepgram differentiates single vs. multilingual modes, while AssemblyAI and ElevenLabs provide automatic detection. Always run a 30-second test before an important meeting.