Natural speech parsing on the phone's neural processor

Spoken instructions are naturally messy and fragmented. Our offline acoustic model isolates intent and temporal entities on-device, turning spoken fragments into confirmed calendar events without streaming audio anywhere.

Acoustic isolation and token streaming#

When speaking into a mobile device, ambient noise, pauses, and false starts complicate standard speech recognition pipelines. Dimlit bypasses traditional remote transcription APIs in favor of a compact, on-device acoustic model compiled directly for hardware neural cores.

TIP

Speaking in complete or rigid sentences is unnecessary. The parser is trained on colloquial phrasing, sentence fragments, and relative time references such as “tomorrow afternoon” or “next Tuesday at three”.

Structured intent representation#

Once raw audio frames are processed into text tokens, an entity extraction layer maps the tokens into strongly typed schema dictionaries:

{
  "intent": "calendar.create_event",
  "confidence": 0.96,
  "entities": {
    "title": "Design review and architecture sync",
    "startDate": "2026-10-18T14:00:00Z",
    "durationMinutes": 60,
    "location": "Design Studio"
  }
}

Parsing stages#

1. Acoustic processing: 16kHz audio samples are converted to log-mel spectrogram features.
2. Phoneme decoding: The neural decoder emits candidate character sequences.
3. Temporal grounding: Relative temporal words are resolved against the device’s local clock.
4. Confirmation card: A visual confirmation prompt renders before committing the calendar event.

By pairing offline speech processing with quantized on-device models, voice interactions remain responsive even when the device is completely disconnected from network access.