--- title: "AI voice dictation: how it learns your specialist vocabulary" description: "Speech recognition, AI correction, correction memory: how modern voice dictation works, and why it stumbles over the technical terms of your profession." url: "https://techvoice.fr/en/blog/how-ai-dictation-works" language: "en" published: "2026-11-09" updated: "2026-11-09" author: "The TechVoice team" --- # AI voice dictation: how it learns your specialist vocabulary > Speech recognition, AI correction, correction memory: how modern voice dictation works, and why it stumbles over the technical terms of your profession. You dictate “hallux valgus correction by Scarf osteotomy”. The screen shows “Puralux valgus correction”. Your microphone is not to blame, and neither is your diction: it comes down to how speech recognition chooses its words. Transcription models learn from vast amounts of everyday speech, in which the terms of your profession are rare. When the sound is ambiguous, the model picks the most likely sequence of words in *its* experience, not yours. Professional voice dictation therefore has to do more than transcribe: it must know your vocabulary, correct without betraying you and remember what you teach it. Here is how that works, layer by layer. 1. **Your voice** — Live dictation or uploaded file 2. **Transcription** — Voxtral, from Mistral AI 3. **Correction** — Using your vocabulary and past corrections 4. **Fidelity check** — Too many changes: the recognised text is kept 5. **Your text** — You review; your edits feed the memory *The TechVoice dictation pipeline. Every step after transcription can fail without any text being lost.* ## Step 1: turning voice into text The first layer is speech recognition, or *speech-to-text*. TechVoice uses Voxtral, Mistral AI’s family of models, in two ways: - **live**, the audio is sent in small chunks and the text appears as you speak; - **on upload**, the model receives the whole recording and has all the context before it writes anything. The second mode makes fewer mistakes. On our French test corpus, the raw text got 25.2% of specialist terms wrong in live mode, against 9.6% on upload, with the latest generation of Voxtral models. Listening right to the end helps resolve ambiguities. Some APIs let you give the model a list of terms to favour, a technique known as *context biasing*. Voxtral’s API accepts one, with limits: 100 terms at most, no spaces within a term, and a feature that Mistral describes as experimental outside English. On our French dictations, this list added nothing once correction had been applied, and even made live dictation worse. So the vocabulary has to do its work elsewhere. ## Step 2: correcting without rewriting Once transcription is complete, a language model rereads the raw text with two pieces of information: your vocabulary and your past corrections. It knows that “Phalux valgus”, in a foot surgery report, is almost certainly “hallux valgus”. On our test bench, this step brings the error rate on specialist terms down from about 25% to about 5% in live dictation. The risk is well known: ask a language model to correct a text and it likes to *improve* it. It smooths a sentence, completes it, reorganises it. For an email, that is a service; for a medical or legal report, it is a danger. Moving a negation is enough to reverse the meaning: “no pulmonary embolism” has nothing in common with “pulmonary embolism”. So TechVoice puts two safeguards in place: 1. **Strict instructions.** The model only corrects recognition errors. It adds, removes and rewords no information, and leaves an ambiguous passage as it is rather than inventing something. 2. **A mechanical check.** After correction, a program compares the corrected text with the recognised text, word by word and **in order**. If the share of changed words exceeds a threshold, the correction is rejected and you receive the raw transcript. Comparing in order is essential: a simple word count would not notice that a negation had moved. The principle is deliberate: better a raw text than a fluent but unfaithful one. And every correction remains visible: one click shows what changed. ## Step 3: a memory that belongs to you The quality of the correction depends on what the model knows about your profession. TechVoice builds that knowledge from three sources, from the most general to the most personal. **Your field.** When you sign up, you enter your field. TechVoice then looks for that field’s terms in the scientific literature of the open archive [HAL](https://hal.science), in the language you dictate in. It is a useful safety net from the very first dictation, without asking anything of you, but it remains generic. **Your documents.** Upload a few sample reports, letters or templates (PDF or Word): TechVoice extracts the terms, then forgets the text. This is the best source to start with. In our measurements, 309 terms taken from a surgeon’s website did better than 3,753 terms drawn from the literature. A lexicon is not better because it is bigger: it is better when it sounds like you. **Your corrections.** This is the heart of the system. When you fix a word, TechVoice records the pair “what was written → what was right”, for example *Phalux → Hallux*. A second check tells genuine recognition errors apart from mere changes of style, so that a rewording on your part does not become a rule. In later dictations, these pairs are passed to the corrector. The effect is measurable. When each dictation in our corpus benefited only from the corrections made on the *other* dictations, the error rate on specialist terms fell from 13.9% to 1.7%. That is better than giving the corrector the exact list of expected terms (2.6%): showing it the precise error and its correction helps more than showing it the right word alone. None of this memory is opaque. In the Memory tab, you can see your vocabulary, your documents and every correction learned, and you can edit or delete them at any time. ## What this changes day to day Here is a real example, checked in the app on a report dictated in French. A surgeon dictates a foot surgery report. Voxtral hears “Puralux valgus”; the corrector suggests “Phalux valgus”, closer but still wrong. The surgeon changes it to “Hallux valgus” and saves. The pair is learned, the term joins their vocabulary, and the same dictation, replayed, now comes out as “Hallux valgus”. That is the promise of dictation that learns: every correction you make works for the ones that follow. And if a layer fails along the way (network outage, unavailable service, rejected correction), the text already transcribed stays put. Nothing you have dictated disappears. ## The limits, honestly - **Our figures are optimistic.** They come from synthetic voices, with no background noise or hesitation. A real voice, a real consulting room and a fast delivery produce more errors. The ranking of the effects, however, holds from one measurement to the next. - **Numbers and units** deserve particular attention when you review: “ten milligrams” and “10 mg” are not written the same way, and a wrong figure cannot be guessed from the context. - **Reviewing remains essential.** Voice dictation, however good, produces a draft of which you remain the author, and for which you remain responsible. To measure what dictation could save you, we have set out [the time-savings calculation, document by document](https://techvoice.fr/en/blog/voice-dictation-time-savings). And to find out more about the transcription model itself, read [our hands-on review of Voxtral](https://techvoice.fr/en/blog/mistral-voxtral-speech-recognition). ## Sources - TechVoice internal measurements, synthetic corpora of reports in orthopaedics and ten specialties, in French, September to November 2026. - Mistral AI, documentation on audio transcription and *context biasing*. - [HAL](https://hal.science), multidisciplinary open archive.