hn.today

Rebuilding my local voice assistant pipeline

blog.platypush.tech127 points23 comments
Screenshot of Rebuilding my local voice assistant pipeline

Platypush’s revised voice-assistant pipeline addresses unreliable wake-word detection and speech-end detection that made its earlier self-hosted setup trigger on noise, miss commands, or keep listening long after a user stopped speaking. It replaces OpenWakeWord with microWakeWord, whose actively maintained models are bundled or can be added from custom manifests. The new plugin preserves the old integration’s configuration and events, making migration straightforward. For speech recognition, the pipeline can use OpenAI’s Whisper service or local Vosk; responses can come from a cloud or compatible local language model, while speech is generated with OpenAI TTS or local Piper.

The key improvement is a shared audio-preprocessing system for detecting when speech ends. It selects Silero VAD by default when available, with WebRTC VAD and an energy-based detector as fallbacks, then combines threshold hysteresis, a half-second sliding-window vote, and adaptive background-noise tracking to reject brief noise bursts. In a synthetic test with babble and burst noise, conversations closed 1.0-1.6 seconds after the final word, compared with 6-10 seconds under the previous setup; Vosk benefits from the same changes. The configuration also pauses wake-word listening during conversations to prevent the assistant reacting to its own speech, and supports hooks for handling commands locally. Results depend on microphones and room acoustics: the author’s Raspberry Pi resamples audio from 44.1 kHz, and users may need to tune detection and noise thresholds.

Read on blog.platypush.tech23 comments on Hacker News

Summary generated by AI from the linked article. hn.today is not affiliated with Hacker News or Y Combinator.

More in AI

The daily digest

Today's best Hacker News stories, summarized and screenshotted, one email a day.