Every word. In context.

I work on multilingual speech-to-text, on-device models, and medical audio—including real-time transcription and understanding. The interesting details are the terminology, the language switches, and what happens when someone corrects themselves mid-sentence.

Three engraved sound-wave ribbons converge into a rose waveform above a tablet and a folded transcript sheet.
From speech to meaning, on the hardware that matters.

Built around the conversation.

I’ve trained audio and voice models. I can help adapt and evaluate a speech pipeline for your languages, vocabulary, devices, and workflow—from a local transcript to structured information a clinician can check.

Multilingual, on-device transcription
Speech-to-text for selected languages and language switching, with latency, memory, and offline operation tested on your target device.
Medical dictation
Transcribe specialist terminology, abbreviations, medication names, and corrections into a draft for clinician review.
Real-time medical understanding
Turn a live conversation into a transcript and draft structured note, with extracted details linked to the relevant audio for checking.

The correction matters as much as the words.

A clinician dictates a note, corrects a detail, and switches language for a name or phrase. The workflow needs to preserve the intended meaning while making it easy to check what was actually said.

Example engagement · Illustrative workflow, not a client result or live model demonstration.

  1. Input

    Authorised audio, the languages in scope, and the agreed note structure.

  2. Model task

    Produce a transcript and draft structured details, retaining links to the audio.

  3. Human decision

    The clinician checks and edits the draft before it becomes part of the record.

What we’d measure.

Meaning, not only word count
Test negation, medication names, dosages, speaker attribution, and self-corrections alongside transcription errors.
Keeping up
Measure time to the first text, revisions while someone speaks, and time to the final reviewable transcript or note.
Real listening conditions
Evaluate the languages and accents in scope, background noise, overlapping speakers, and sustained performance on the target device.

We scope transcription, understanding, and note generation separately. Draft clinical content requires clinician review; processing location, recording consent, retention, and access are agreed with your team.

Let’s test the audio you actually hear.

My experience includes training audio and voice models. For a new engagement, I start with representative recordings and the hardware constraints, then compare approaches before committing to a larger build.

See the model feasibility sprint

Start with one useful test.

Bring one task, a few representative examples, and the constraints that matter. We’ll agree what to compare before starting.

Model feasibility sprint

A$5,500

One task, a baseline, and one alternative. You get the comparison, the important failure cases, and a recommendation on what to build or keep.

Need an initial second opinion? The A$1,250 production review is a smaller place to start. Larger builds and production integration are scoped separately.

A few useful questions.

Will everything run on-device?

We test that rather than assume it. Speech recognition, speaker handling, and note generation may have different hardware needs. We agree which parts must stay on-device and measure the complete proposed pipeline.

Which languages can we support?

We select the languages, accents, and language-switching patterns during scoping, based on the available models and representative evaluation audio. Support for a language does not establish medical accuracy in that language.

Does real-time mean the text never changes?

No. Streaming transcripts can revise earlier words as context arrives. We measure those revisions and the finalisation delay, so the interface distinguishes provisional text from the version ready for review.

Bring the awkward examples.

Tell me about the task and where the current approach falls short. A description is plenty for the first conversation; we’ll agree how to share any sensitive material.