What it's for

Every speech model has words it keeps getting wrong: names, products, technical terms. Your dictionary remembers those mistakes and fixes them in each new dictation, before the text reaches you.

How an entry works

An entry pairs what the speech model writes with what you might have meant. Say it often writes “cloud” when you said “Claude”. But sometimes you really mean a cloud. So the entry keeps both meanings:

In “Ask cloud to draft the reply”, Entune writes Claude. In “The backups are in the cloud”, it leaves cloud alone. It chooses from the surrounding words, which is why each meaning has a short description: write what the thing is and what it goes with.

One entry can list several spellings the model writes (“cloud”, “clawed”) and several meanings. Each heard spelling is linked to the meanings it can stand for.

Getting suggestions

Entune can build the dictionary for you. A suggestion model, the language model you choose with its key in Settings, reads what your speech model wrote and suggests entries for the words it gets wrong. For the entries those transcripts show, it also suggests clearer meanings, more heard spellings, and removals.

Get suggestions is the simple path: it reads up to 300 of your speech model's newest transcripts not yet used by suggestions you applied; a run you discard or stop uses none.

Reusing earlier transcripts

Each transcript is read once: after you apply a run's suggestions, the next Get suggestions reads only what you dictated since, so you don't pay twice for the same text. Reuse earlier transcripts reads them again. It shows every transcript of this speech model on a timeline, oldest on the left, with the ones used before marked under the line; drag the handles to choose how much to read.

Reuse them when the earlier suggestions no longer stand for what those transcripts can teach: you cleared or rebuilt your dictionary, you chose a stronger suggestion model or a higher reasoning effort, an update improved how Entune suggests, or an earlier run's suggestions were poor. Reading again costs tokens again. Entries already in your dictionary are shown to the suggestion model beside the transcripts they occur in, so it improves them instead of adding them twice.

Learning from audio

The ⋯ menu learns from audio instead of transcripts, from one of three sources:

Your other models' recordings are recordings already in Entune that this speech model hasn't transcribed. Another dictation app: Import copies the audio that app keeps on this Mac into Entune's data folder; its transcripts are never copied and its files stay as they are. Importing again adds only what is new. An audio folder: choose a folder of recordings of you talking, such as meetings or voice notes, in WAV, MP3, M4A, FLAC, OGG or WebM; each file is copied once.

The hours and minutes are recorded audio: the length of the recordings you choose, not how long the run takes. Entune first transcribes that audio with your speech model, which a cloud provider charges for, then the suggestion model reads the new transcripts. The estimate under the timeline shows both steps. The timeline runs oldest on the left; drag the handles to choose how much to use, and each edge snaps to whole recordings. An hour or two of recent audio gives quick results.

Time and settings

Text goes to the suggestion model in parts, one after another, because each part builds on the entries found so far. With audio, parts start as soon as enough is transcribed, while the rest is still transcribing. Cloud speech models transcribe several recordings at a time; a local model takes one at a time.

Reasoning effort is the level the suggestion model's provider names, from minimal to xhigh. Higher effort reasons longer and usually finds more; high is the default. A provider without a level maps it to one it has.

The estimate uses what Entune has measured: your speech model's transcription times, and the time per part of earlier runs with the same suggestion model and effort. A combination not used before says so.

Sometimes a reply starts and then sends only blank space. Entune stops it within a minute or two and tries that part again, three attempts in all. A ChatGPT plan also ends any single reply at about 15 minutes. If a part fails all its attempts, or runs past a time or length limit, Entune stops and keeps the finished parts. Choose a lower reasoning effort, or less audio, and press Continue: it picks up where it stopped. Transcripts made so far are kept for Continue while Entune stays open.

Reviewing suggestions

Suggestions open in a panel beside your dictionary, grouped into new entries, changes and removals. Edit any of them, leave out the ones you don't want with ×, and apply the rest in one step. Nothing is saved until you apply. Dictation keeps working the whole time; only editing the dictionary waits until you apply or discard.

Each speech model makes its own mistakes, so suggestions are kept for the speech model chosen when the run starts.

Pinned and learned entries

Learned entries are used only with the speech model they were learned for. Pin an entry to use it with every speech model; pinned entries are yours and suggestions can't remove them. Pinning doesn't make an entry win: Entune still reads the sentence.

Adding an entry yourself

Choose Add and fill in one line: how it's written, what the speech model hears (several, separated by commas) and what it is. For several meanings, a spelling kept as written, or always replacing, open the full editor from the link under that line. Edit on an entry opens the same editor.