Skip to main content
Klaus uses OpenAI’s text-to-speech for output and Moonshine for local speech-to-text. Both are highly configurable.

​
Text-to-Speech (TTS)

​
Voice Selection

​
voice
string
default:"cedar"
The OpenAI TTS voice model to use for spoken responses.Available voices:
  • alloy
  • ash
  • ballad
  • cedar (default)
  • coral
  • echo
  • fable
  • marin
  • nova
  • onyx
  • sage
  • shimmer
  • verse
Listen to voice samples at OpenAI’s TTS documentation to find your preferred voice.

​
Playback Speed

​
tts_speed
float
default:"1.0"
Playback speed multiplier for TTS audio.Valid range: 0.25 to 4.0
  • 0.25 = 4× slower (very slow)
  • 1.0 = normal speed (default)
  • 2.0 = 2× faster
  • 4.0 = 4× faster (maximum)
Values outside the 0.25-4.0 range may produce distorted or unintelligible audio.

​
TTS Model

Klaus uses OpenAI’s gpt-4o-mini-tts model, which is hardcoded and cannot be changed via configuration. Cost: Approximately $0.015 per minute of generated audio.

​
Voice Instructions

Klaus sends the following instructions to the TTS API to optimize voice output:
“Speak at a natural conversational pace, not slow or deliberate. You are a sharp colleague giving a quick answer across a desk. Be direct and matter-of-fact, not performative. No vocal fry, no uptalk.”
These instructions are hardcoded and ensure consistent, professional voice output across all voice models.

​
Speech-to-Text (STT)

Klaus uses Moonshine Voice, a local on-device STT model that runs entirely on your machine with no API calls or costs.

​
Model Size

​
stt_moonshine_model
string
default:"medium"
Moonshine model size. Larger models are more accurate but slower.Available models:
ModelSizeLatencyAccuracy
tinySmall~100msGood
smallMedium~200msBetter
mediumLarge~300msBest (default)
The medium model (default) provides the best balance of accuracy and speed for most users. The model is downloaded automatically on first use.

​
Language

​
stt_moonshine_language
string
default:"en"
Language code for Moonshine transcription.Default: en (English)
See the Moonshine documentation for a full list of supported language codes.

​
Input Mode

​
input_mode
string
default:"voice_activation"
Default input mode on startup.Options:
  • voice_activation: Klaus automatically detects when you’re speaking (default)
  • push_to_talk: Hold the push-to-talk key to record
You can toggle between modes at runtime using the toggle key (default: F3 on Windows, § on macOS).

​
Example Configuration

config.toml
# Use the 'nova' voice at slightly faster speed
voice = "nova"
tts_speed = 1.2

# Start in push-to-talk mode
input_mode = "push_to_talk"

# Use small Moonshine model for faster transcription
stt_moonshine_model = "small"
stt_moonshine_language = "en"

​
Advanced: TTS Streaming

Klaus uses sentence-level streaming for low-latency responses:
  1. Claude’s response is split into sentences as it streams
  2. Each sentence is sent to OpenAI TTS immediately (max 4000 chars per call)
  3. Audio playback starts on the first chunk
  4. Remaining chunks play seamlessly as they’re generated
Result: You hear the first sentence in 2-3 seconds, well before the full response is complete.

​
Platform Optimizations

  • macOS: Uses high latency mode to prevent CoreAudio crackling
  • All platforms: Reuses a single persistent audio output stream across all chunks to avoid device initialization delays
  • VAD suspension: The microphone stream is suspended during TTS playback to free the audio device

​
Latency Breakdown

Typical end-to-end latency from question to first spoken word:
StageLatency
VAD detection + silence timeout0.5-1.5s
Moonshine STT (medium model)~300ms
Claude vision + reasoning (first chunk)1-2s
OpenAI TTS (first sentence)0.5-1s
Total2-4 seconds
Subsequent sentences stream with minimal additional latency, creating a natural conversational flow.

​
Model Downloads

The Moonshine model is downloaded automatically on first use:
  • tiny: ~80 MB
  • small: ~160 MB
  • medium: ~245 MB
Models are cached locally and only downloaded once.

​
Troubleshooting

​
TTS Voice Not Working

  1. Check API key: Verify OPENAI_API_KEY is set correctly
  2. Check voice name: Ensure the voice name matches one of the available options (case-sensitive)
  3. Check logs: Look for OpenAI API errors in Klaus’s console output

​
STT Transcription Inaccurate

  1. Try a larger model: Switch from tiny to medium for better accuracy
  2. Check microphone: Ensure your mic is selected correctly in Settings
  3. Reduce background noise: Moonshine works best in quiet environments
  4. Adjust VAD sensitivity: See Advanced Settings

​
TTS Playback Too Fast/Slow

  • Adjust tts_speed in config.toml
  • Valid range is 0.25 to 4.0
  • Recommended range: 0.8 to 1.5 for natural-sounding speech

Build docs developers (and LLMs) love

Get started for freeTalk to us