Speech to text without invented words

Transcribe

Turn speech into timed text locally, using an acoustic system built to emit only words supported by the audio.

For interviews, meetings, accessibility, and media work where a fluent invention is worse than an honest gap.

Current availability

The Rust engine and CLI work with prepared models. There is no public installer or browser service, and model tests remain hand-run.

Standalone

Why use it

A smaller path from intent to outcome.

  1. 01

    Keep audio local

    CPU inference does not require sending a recording to a hosted model.

  2. 02

    Retain the timing

    Word and token timestamps keep the transcript connected to the recording.

  3. 03

    Prefer a gap to a guess

    The non-generative pipeline is designed around acoustic evidence.

Development record

What is real now—and what still has to be earned.

The badge names this product’s place in the portfolio, not public availability. This record separates working parts from missing product proof.

In the code today

  • Pure-Rust VAD, mel frontend, FastConformer, CTC, and TDT inference
  • Three model tiers with recorded word-error measurements
  • Timed word and token JSON output on CPU, with a CUDA path

Before broader release

  • Reproducible model-bearing CI instead of hand-run gates
  • Installable releases and a local accessibility-first desktop surface

Place in Argand

Transcribe is a standalone accessibility and media tool, not an LLM feature.

Continue exploring

Meet Account-optional, end-to-end encrypted calls with local recording. Archive Self-hostable research archives with evidence capture, timelines, maps, and citations. Originality Check text against web sources and inspect matching passages without storing it.