Clean voice transcription for professionals — making technical vocabulary survive dictation

Voice dictation works beautifully — right up until you use the vocabulary of your profession. A doctor dictates "started him on metoprolol fifty milligrams" and gets "started him on met a pro lol fifty milligrams." A litigator says "voir dire" and the transcript reads "vwa deer." An engineer says "Kubernetes" and gets "cooper netties."

For casual users this is a funny quirk. For professionals it's disqualifying: if you have to stop and fix every technical term, the speed advantage of dictation evaporates. This post explains why it happens and what a real fix looks like — not "train your voice profile for 30 minutes," but clean voice transcription that handles your field's vocabulary out of the box.

Why professional vocabulary breaks ordinary dictation

Speech-to-text models learn from enormous amounts of mostly-everyday speech. That training makes them excellent at common words and statistically biased against rare ones. When the audio is ambiguous — and real-world audio always is — the model picks the likelier word. "Metoprolol" is rare in general speech; some mush of common syllables is, statistically, the "safer" guess.

This is why the failure feels so specific to professional work. Your everyday sentences transcribe perfectly; the moment you use the vocabulary that makes your dictation professionally useful, the model swaps it for noise. The accuracy you experience isn't the model's average accuracy — it's the accuracy on exactly the words you care about most.

What doesn't fix it

• A better microphone. The model isn't mishearing — it's mis-guessing. Cleaner audio narrows but doesn't close the gap.
• Talking slower. Same reason. The model's vocabulary prior, not your articulation, is the bottleneck.
• Post-hoc find-and-replace. A replacement table that maps "met a pro lol" → "metoprolol" only works when the error is identical every time. Speech recognition errors aren't — the same word gets mangled five different ways on five different days.

What actually fixes it: three layers

Layer 1 — tell the transcription what to expect

Modern transcription models accept a vocabulary hint: a list of terms the audio is likely to contain. With the hint in place, the model's guess flips — it now expects "metoprolol," so the ambiguous audio resolves to the right word. This happens before any text exists, which makes it the highest-leverage fix. Pithflow sends this hint on every single dictation, built from your dictionary and your enabled term packs.

Layer 2 — repair near-misses during cleanup

Pithflow runs every transcript through an AI cleanup pass (the same one that strips filler words and fixes punctuation). That pass also gets your vocabulary as a glossary, with one job: if a phonetically-similar mangled form appears, restore the correct term — and never substitute when the match isn't clear. This catches what layer 1 misses.

Layer 3 — deterministic replacement as a backstop

The classic dictionary find-and-replace still runs last, for the handful of errors that are consistent. Three layers, one goal: the term you said is the term that lands in your document.

Specialty term packs: the vocabulary, pre-built

The layers above need vocabulary to work with. You could hand-enter hundreds of terms — or enable a curated pack for your field in one click. Pithflow ships eleven, nearly 2,000 terms in total:

• Medical (English) — top-prescribed drugs, clinical exam vocabulary, diagnoses, anatomy, procedures. The "metoprolol problem," solved at the source.
• Médico (Español) — vocabulario clínico y fármacos for Spanish-speaking clinicians, tuned for Mexican and LATAM usage.
• Software & engineering — Kubernetes, PostgreSQL, OAuth, idempotency, and the rest of the vocabulary that turns into soup in generic dictation.
• Legal (English) — voir dire, res judicata, estoppel, interrogatories; litigation and contract vocabulary.
• Legal (Español) — amparo, juicio de nulidad, derecho procesal y notarial mexicano.
• Finance & accounting — EBITDA, amortization, 10-K, accruals, the acronym-dense language of reporting.
• Science & research — PCR, CRISPR-Cas9, spectrophotometry, p-values; lab and publishing vocabulary.
• Mental health & therapy — CBT, EMDR, DSM-5-TR diagnoses, treatment-plan and progress-note language.
• Education & teaching — IEP, 504 plan, formative assessment, phonemic awareness; the vocabulary of school paperwork.
• Construction & trades — rebar schedules, change orders, HVAC, GFCI; field notes and permit language.
• Marketing & sales — CTR, ROAS, attribution, programmatic; the metric soup of campaign reporting.

New users pick their field at signup and the right pack is on from the first dictation. Existing users toggle packs in the Dictionary tab. Packs stack with your personal dictionary — your terms always win.

Bilingual professionals: the Spanish case

A large share of professional dictation in the Americas is bilingual. A Mexican physician's chart note might read: "paciente con hipertensión arterial, started him on metoprolol, control en dos semanas." Two languages, technical vocabulary in both, switched mid-sentence.

Pithflow auto-detects the language per dictation and applies the matching pack vocabulary — enable both Medical (English) and Médico (Español) and the code-switched sentence above transcribes correctly in both halves. No language toggle, no separate modes.

When a term still slips through

Packs cover each field's high-yield vocabulary, not every regional brand name, niche subspecialty term, or colleague's surname. For those:

1. Personal dictionary — add the term once; from then on it's part of your recognition vocabulary on every dictation.
2. Fix a word — in your History, one click on any past dictation: type what it got wrong, what you actually said, save. The correction is permanent for your account, and recurring fixes across users tell us which terms to promote into the packs for everyone.

This is the part most dictation tools skip: a feedback loop. The packs aren't a static list someone guessed at — they grow from the corrections professionals actually make.

What "clean transcription" means in practice

Put the vocabulary work together with the cleanup pass and the output changes category. Raw dictation:

"um so the patient is a fifty four year old male uh with two week history of chest pain started him on met a pro lol and uh told him to come back in two weeks"

With vocabulary biasing + cleanup:

"The patient is a 54-year-old male with a two-week history of chest pain. Started him on metoprolol; follow-up in two weeks."

No filler, correct punctuation, correct drug name. That's the bar for professional use: the transcript is done when it lands, not a rough draft you then repair.

A note for healthcare readers

Better transcription doesn't change compliance posture. Pithflow's current tiers are not HIPAA-eligible — an Enterprise tier with BAA is planned for Q4 2026. Until then, don't dictate PHI; the packs are just as useful for the non-PHI writing (referral letters with names removed, teaching notes, research drafts) that fills the rest of a clinical day.

Pithflow is voice dictation with AI cleanup for Windows + Mac. Download free — 2,000 words/week, no credit card. See also: Specialty term packs · Pithflow for doctors.