A Brisbane GP I know runs a billing experiment every quarter. She picks a fortnight and tracks every minute she spends on documentation — letter dictation, GP-to-specialist referrals, hospital discharge summaries, the running commentary on each patient that becomes the consult note. Last quarter the total came to 11 hours and 40 minutes across two weeks, which is essentially a full clinical day per fortnight burned on typing.

That's not unusual. The RACGP estimates Australian GPs spend 1.5–2 hours per clinical day on documentation, with the higher end common in older or comorbid patient cohorts. Multiply across the ~32,000 practising GPs in Australia and you get a workforce-level inefficiency that's been visible for a decade and obvious for the last three.

The dictation market has tried to fix this twice. First, traditional typing services — outsourced typing pools in Australia or the Philippines that take audio recordings and return typed letters in 2–8 hours. Second, voice-to-text software like Dragon Medical, which transcribes locally but requires extensive per-clinician training.

The third option — generic AI dictation built on Whisper-class models — is the one most practices are now testing. This piece breaks down how it compares on the four things GPs actually care about: accuracy on clinical vocabulary, speed, cost, and integration with the existing PMS.

What "AI dictation" means in 2025

The category has fragmented. There's now a clear distinction between:

Real-time dictation — like Dragon Medical One, which converts speech to text in your active text field as you speak. Useful for consult notes typed during the consult.
Asynchronous transcription — services like speechtotext.au where you upload an audio file (or recording from your phone) and receive a transcript back in 30–90 seconds. Useful for letters, referrals, summaries, and anything you'd otherwise hand to a typist.
Ambient AI scribes — Heidi, Lyrebird, Suki and similar tools that listen to the consult itself and generate the consult note structured to SOAP format. A separate category we'll cover elsewhere.

This article focuses on asynchronous transcription, because it's the cleanest replacement for traditional typing services and because the cost-benefit comparison is the most straightforward.

Accuracy on Australian clinical vocabulary

Generic Whisper-large-v3 is trained on a broad multilingual corpus that includes a meaningful chunk of medical content — published podcasts, medical YouTube channels, clinical lectures. It performs better on clinical terms than most GPs expect on first test.

We benchmarked a representative sample of Australian clinical dictation against three categories of terms:

Medication names (PBS-listed): Whisper got 96% of common medication names correct on first pass. Failures clustered on rarer biologics (adalimumab came through as "Adam Lumab") and PBS-specific brand variants (Aspen Levothyroxine was sometimes split). For the top 200 most-prescribed medications in Australia, accuracy is essentially perfect.

Australian-specific health-system terminology: Better than expected. "MBS item 36" came through correctly. "ATSI" was recognised. "GP Management Plan" and "Team Care Arrangement" were both correct. The model has clearly seen Medicare-adjacent vocabulary.

Specialist subspecialty terms: Where the model struggles. Cardiothoracic vocabulary ("posterior tibial artery", "atrioventricular nodal reentrant tachycardia") gets 80–85% right. Oncology drug protocols (FOLFIRINOX, AC-T) come through but specific dosing schedules sometimes need correction.

Aussie patient-history idioms: Surprisingly good. "Patient presents with a sore arvo" wasn't fabricated — but if a patient says "I had a fall this arvo," the model gets it right.

The headline takeaway: for the 80% of GP dictation that's letters, referrals, and standard consult content, AI dictation produces a document that needs minimal correction. For sub-specialty referrals heavy in obscure pharmaceutical chemistry, you'll still want a 30-second review pass.

Speed comparison

A 4-minute dictated letter through a traditional typing service comes back in 4–8 business hours. The actual range depends on the service:

Australian-based services (e.g. Medical Typing, Audio Typing Australia): 4–6 hours during business days. After-hours dictation queues for morning.
Offshore services (Philippines-based): 2–4 hours with 24-hour coverage. Some lag at peak times.
Hybrid services: a 24-hour turn-around for "standard" jobs and 1-hour for "rush" at 2× the rate.

AI dictation through speechtotext.au returns the same 4-minute letter in about 30 seconds. The accuracy isn't quite human-typist level (typists understand context; AI doesn't), but the speed difference fundamentally changes the workflow. When dictation comes back during the consult, you can sign and send it before the patient leaves. When it comes back tomorrow, you have to remember to action it.

Several practices we work with report this is the single biggest unexpected benefit: not the cost reduction, the flow of having documentation finalised within the same consult block.

Cost comparison

The cost piece is where the gap is largest.

Traditional typing services charge between $1.80 and $2.80 per audio minute, depending on turnaround and volume commitment. A typical GP dictates roughly 15 minutes of audio per clinical day across letters and notes. That's $27–$42 per day, or $135–$210 per clinical week, or roughly $7,000–$11,000 per GP per year.

A practice with 6 FTE GPs running typing services is therefore spending $42,000–$66,000 per year on dictation.

AI dictation through speechtotext.au at the Business plan ($49/month, 4,000 minutes) costs $588 per GP per year and covers their entire dictation volume with substantial headroom. For 6 GPs, that's $3,528 per year — a 90%+ cost reduction.

The smaller observation worth flagging: the typing-service cost is variable (you pay per minute), while AI dictation is fixed. Practices with seasonal patient-load fluctuations end up with predictable monthly software cost vs surprisingly large bills in flu season under typing services.

The Medicare item-number question

The most common compliance question we get is whether AI dictation affects what GPs can bill.

The short answer: no. MBS items 23, 36, 44, 5040 (standard, long, prolonged, after-hours consultations) and the chronic disease management items (721, 723, 732) don't specify how documentation is produced. Whether you type the note yourself, dictate to a typist, or use AI dictation, the MBS billing rules don't change.

The longer answer requires more care. AHPRA's Code of Conduct (section 4.4) requires that medical records be "clear, accurate, contemporaneous, and complete". Three implications:

Accurate: You're responsible for what's in the note. If AI dictation produces a transcript with an error and you sign it without reading, you've signed an inaccurate record. This is the same standard as typing services — you've always been responsible for reviewing what comes back.

Contemporaneous: Notes should be produced "at or near the time of the event". AI dictation is actually better here than typing services, because it returns the document within the consult block rather than 6 hours later. The contemporaneous standard is easier to meet, not harder.

Privacy of audio: Section 11.2 of the Privacy Act and the APPs require that "reasonable steps" be taken to protect health information. Audio files containing patient details fall squarely under "sensitive information" (APP 3.3). Practical effect: pick a service whose audio handling you can audit. speechtotext.au processes audio in memory and discards it after transcription, with no on-disk storage. Services that store audio for "model improvement" are a flag.

We have a medical use-case page that summarises the privacy posture if it helps with practice-policy compliance.

Integration with Australian PMS systems

The blocker for most practices isn't the transcription quality — it's how the transcript gets into the patient record.

MedicalDirector / MD3: No direct API. Practice workflow is typically: dictate on phone → upload to speechtotext.au → copy the transcript text → paste into the consult note. Two clicks once the workflow is set up. Some practices use a Windows-side clipboard-watcher (PhraseExpander) to make the paste a single hotkey.

Best Practice (Bp Premier): Same pattern. No published API. Copy-paste workflow.

Halaxy: Has a free-text consult-note field that accepts paste, and a separate "Letters" section that accepts paste-or-rich-text. AI dictation paste works in both.

Genie: Same pattern. Free-text fields throughout, paste works fine.

Communicare (used in Aboriginal Community Controlled Health Services): Same paste pattern.

The honest assessment: no PMS in Australia has a first-class AI-dictation integration as of late 2025. Every practice we work with has settled on the same pattern — dictate on phone, transcribe through speechtotext.au, paste into the PMS. The friction is acceptable because it's just one paste, and it's faster than any alternative.

What still needs a human

A short list of things AI dictation doesn't yet handle well:

Hand-drawn diagrams (musculoskeletal injury notation, fracture descriptions). Use BetterConsult or a separate sketch tool; the transcript handles the text.
Numeric prescription instructions with complex dosing ("take 1 tablet daily for 3 days, then 2 tablets daily for 4 days"). Whisper sometimes substitutes numbers; always review.
Patient names with non-English spellings. The model produces a phonetic best-guess. A 5-second review catches this.
Death certificates and medico-legal letters. Same accuracy review you'd apply to a typing service, but the stakes are higher; don't sign without reading.

The pattern is the same as for any AI assistance in clinical practice: it makes the first draft instant, but the clinician remains accountable for what gets signed.

Practical workflow — the 5-minute setup

For a GP wanting to test AI dictation alongside their existing typing service:

Day 1: Pick a slow afternoon. Dictate a letter on your phone (Voice Memos on iPhone, Easy Voice Recorder on Android). Upload at speechtotext.au. Compare the result to what you'd expect from your typist.
Day 2–5: Use AI for routine letters; keep complex referrals on the typing service.
Week 2: If accuracy is acceptable, move 80% of dictation to AI. Keep typing service for high-complexity work.
Month 2: Most practices have moved to 100% AI by this point.

The free tier covers about 2–3 letters per month — enough to test, not enough to run a practice. Pro at $19/month covers a part-time GP. Business at $49/month covers a full-time clinical week with headroom.

FAQs

Is AI dictation legal for use in Australian general practice?

Yes. The MBS doesn't specify how documentation is produced, and AHPRA's Code of Conduct standards (clear, accurate, contemporaneous) are easier to meet with AI dictation than with delayed typing services. The privacy of audio handling needs to comply with APP 3.3 — use a service whose audio handling you can audit.

Will Whisper understand medical terminology?

For the top 95% of clinical vocabulary used in Australian general practice, yes. Sub-specialty pharmaceutical terms sometimes need correction. The model handles Australian-specific terminology (MBS items, ATSI, GPMP, TCA) correctly.

How does the cost compare to my typist?

At typical Australian rates of $2/audio minute, an AI dictation service costs roughly 10% of a typing service for equivalent volume. A 6-GP practice typically saves $40k–$60k per year.

Where does the audio go?

With speechtotext.au, audio is processed in memory and discarded after transcription — no on-disk storage. Verify this for any service you evaluate; some store audio for "model improvement", which is a privacy red flag for clinical use.

Can I integrate with MD3 or Best Practice?

Not directly through an API. The workflow is dictate → transcribe → paste into the consult note. One click once set up, faster than any alternative for the foreseeable future.

What about ambient AI scribes (Heidi, Lyrebird)?

Different category. Ambient scribes listen to the consult itself and generate structured SOAP notes. Useful for consult notes; not a replacement for letter dictation. Most practices end up using both — ambient for notes, async transcription for letters.

---

Start a 14-day Pro trial or [email us](mailto:hello@speechtotext.au) if you'd like a practice-wide demo with our medical use-case team.