Voice to text transcription software in 2026: the two axes every shortlist forgets

Almost all of the modern options stream raw PCM audio (typically 16-bit, 16 kHz, mono or stereo) over a long-lived connection to a hosted speech-to-text model and receive interim and final segments back. The model returns per-word timestamps, per-word confidence scores, speaker labels (diarization), and an `is final` flag so the calling app can decide when to commit text. Older Windows-era apps like Dragon still ship with on-device acoustic models, but the cloud-streamed shape is the default for any 2024 or newer product. The differentiator between apps is mostly which model they call (Deepgram Nova, OpenAI Whisper, Google STT, AssemblyAI, Apple SpeechAnalyzer, Apple Speech, WhisperKit on-device) and how the resulting text is post-processed before display.

On clean, single-speaker English in a quiet room with a decent USB or laptop mic, every name on the popular shortlists (Otter, Sonix, Rev, Wispr Flow, Apple Dictation, Deepgram Nova-3, OpenAI Whisper Large) lands in the 93-97 percent word accuracy range. On noisy audio, multi-speaker, or audio with technical proper nouns, the spread widens fast. The vendors that publish their own benchmarks (Deepgram, AssemblyAI, OpenAI) are not lying about their numbers; they just measured on the friendliest possible audio. For your own audio the only reliable benchmark is to run a 20-minute sample through three candidates and count substitution errors yourself.

Dictation software targets a focused text field in a foreground app and types into it as you speak. Transcription software ingests pre-recorded or streaming audio and produces a saved text artifact, usually with timestamps. The boundary blurs in 2026: Wispr Flow is sold as dictation but emits a polished transcript, Otter is sold as transcription but ships a desktop dictation feature. The boundary that still matters is whether the output is destined for a human reader or for another machine to act on. That is the axis that changes the engineering, not the marketing label.

Yes for the model layer (OpenAI Whisper, NVIDIA NeMo Parakeet, WhisperKit on Apple Silicon all run locally with respectable accuracy), and yes for the app layer if you accept that the consumer-grade product polish is shallower than the SaaS options. The advantage is not price; cloud Whisper API is cheap, and Apple Dictation is free. The advantage is auditability. With an open-source app you can read the vocabulary list, the spoken-form replacement table, the parameters passed to whichever model is being called, and the post-processing logic. With Sonix or Otter you cannot. If your transcripts feed an action layer or sit inside a regulated workflow, that visibility becomes load-bearing.

When a person dictates a URL, they say `fazm dot ai slash download`. The base model transcribes that literally. If the transcript is for a human, the human silently reads `fazm.ai/download` and moves on. If the transcript is going to drive a browser, the agent types `fazm dot ai slash download` into the address bar and fails. Spoken-form replacement rules collapse spoken phrases into their written equivalents at the transcription layer, before any downstream component sees the text. Deepgram exposes this as a `replace=` URL parameter; OpenAI Whisper has prompt seeding that performs a similar function less precisely. Fazm ships 18 default replacements (ten domain TLDs, the at sign, seven source-file extensions). The list is in the public repository and is editable per user.

Vocabulary biasing pre-loads the model with terms it would otherwise treat as unfamiliar word salad. The most common failures are technical proper nouns (`Anthropic`, `Supabase`, `MCP`, `ACP`, `Claude`, `Sonnet`, `PostHog`), product names that collide with common English (`Sonnet` becoming the Shakespeare reference, `Stripe` becoming the verb), and names of people on your team. Deepgram Nova-3 calls these `keyterm` and accepts up to 500 per request, but effectiveness drops sharply past about 30 because the bias collides with itself. Fazm seeds 19 system terms and lets the user add their own. The seeded list is visible in the source as `systemVocabulary: [String]`. If your transcripts include any technical jargon at all, this single feature usually moves accuracy more than picking a different vendor.

Apple Dictation (built into macOS and iOS) and any WhisperKit-based app run on-device with no audio leaving the machine. Apple Voice Control is also fully on-device. The trade is that on-device options have a smaller vocabulary surface than cloud Nova-3 or Whisper Large and tend to be worse on technical jargon. If privacy is the gate but you still want hosted-grade accuracy, look for a product that lets you bring your own API key for an EU-resident inference endpoint, or one that supports a custom API endpoint so you can route through a privacy-respecting proxy.

If the work is purely typing into apps, Wispr Flow and Apple Dictation cover most of it; Wispr Flow is easier to live with day to day on macOS in 2026. If the work is `do this thing for me` rather than `type this out`, you have crossed into desktop agent territory, and the right product is one that pairs transcription with computer control (Fazm, or a chained setup of dictation app plus separate agent). The category boundary matters because dictation apps stop after the text is typed; an agent runs the actions implied by the text.

Fazm transcribes voice into intent, then drives the rest of macOS through accessibility APIs to carry that intent out. The transcription layer is tuned for action: keyterm vocabulary biasing for the names the user mentions most, spoken-form replacement so URLs and email addresses become typeable strings, multichannel capture so the user's mic and the system audio land on separate channels with separate speaker labels. The repository is `github.com/mediar-ai/fazm` and the relevant Swift file is `Desktop/Sources/TranscriptionService.swift`. The vocabulary seed is in `Desktop/Sources/DeletedTypeStubs.swift` around line 657.

Three concrete tests. First, record a one-minute clip of yourself speaking the way you actually speak, jargon included, and run it through every candidate. Compare substitution errors on the proper nouns. Second, dictate a short email that includes a URL, an email address, and a code-style token like `app.json`. Verify whether the spoken-form rewrites happened. Third, look at whether you can read the vocabulary and replacement rules without filing a support ticket. The third test eliminates most of the popular shortlist and surfaces the open-source candidates plus the few SaaS products with a real custom-vocabulary settings panel.