Loquela
Research

Transcription and AI rewriting are different products

Voice transcription tries to preserve what a person said. AI cleanup interprets and changes it. That product choice affects latency, privacy, and authorship.

2 min readResearch checked 2026-07-26

Speech transcription and AI rewriting can appear in the same keyboard, but they solve different problems.

Transcription tries to turn spoken words into text while preserving the speaker's wording. AI cleanup reads that result and decides how to restate it. The second system may remove filler, change tone, restructure a sentence, or replace words that it considers less suitable.

Loquela stops at transcription.

Why the distinction matters

Authorship

A transcription error is a failure to capture what the speaker said. A rewrite can be fluent and still no longer be the speaker's sentence.

That may be useful for people who explicitly want a writing assistant. Loquela is built for people who want their own words entered into the field.

Processing scope

A cleanup model must process the text or utterance it changes. This is an architectural observation, not a claim that a particular keyboard reads every unrelated field on a device.

For example, Wispr Flow's current public site advertises voice dictation with AI commands and auto-edits. Those functions need the submitted utterance and its transcription in order to decide what to change. See Wispr Flow's product description.

Loquela's local path gives the recording to the selected speech model and commits the resulting text. It does not send the transcript through a second general writing model.

Latency and failure modes

Every additional interpretation stage adds work and another place where meaning can change. A cleanup system may produce polished text after seeing the whole thought. A direct transcription system can return partial or final words without waiting for a separate rewrite.

The tradeoff is real. Loquela will not silently turn fragments into polished prose or infer an unstated tone. Users edit the result themselves with normal keyboard controls.

Privacy

Local transcription can run without sending audio or text to a server after a model is installed. Optional Remote transcription sends only recordings from sessions the user starts. In either mode, Loquela does not add an AI rewriting pass.

This does not make every local model accurate or every remote product unsafe. It defines which data must be processed to provide the selected feature.

The product decision

Loquela treats AI as a tool that should be invoked for a specific task. The keyboard does not place a general interpreter between the user and every sentence.

That decision rules out some popular cleanup features. It also gives the product a clear boundary: voice produces text, then the user remains the writer.