A transcript that doesn’t know who is talking is only half a transcript. “We agreed to move the deadline” is a very different sentence when you know whether the CEO said it or the intern did. Speaker diarization—the technology that separates a recording into “who spoke when”—is what turns a wall of words into a record of a conversation.
Most tools that offer it have one thing in common: they do it on someone else’s servers. TranscribeX takes the harder, better road. It runs a research-grade diarization pipeline directly on your Mac, accelerated by Apple silicon’s Neural Engine, and pairs it with Voice Profiles that recognize the people you talk to regularly—so your transcripts say Maya and Daniel, not Speaker 1 and Speaker 2.
What speaker diarization actually does
Diarization answers a deceptively hard question: in a single audio stream where voices interrupt, overlap, and trail off, which stretches of sound belong to the same person? A diarization system listens to the acoustic texture of the recording—pitch, timbre, cadence—divides it into speech regions, computes a compact “voice fingerprint” for each region, and clusters the fingerprints so that every cluster is one speaker.
Done well, it changes what a transcript is for. Meeting minutes can attribute decisions and action items to owners. An interview separates the question from the answer. A podcast transcript becomes readable dialogue instead of a monologue mash. And search gets sharper: you can find not just what was said about the budget, but what Priya said about the budget.
How TranscribeX separates voices
When you run diarization on a transcribed recording—or when it runs automatically after a meeting—TranscribeX executes a four-stage pipeline on your Mac:
Under the hood this is the open pyannote community-1 pipeline running as native Core ML models: powerset segmentation to find who-is-speaking regions (including overlapping speech), WeSpeaker embeddings to fingerprint each voice, and VBx clustering—a variational Bayes refinement over an agglomerative first pass—to decide how many distinct speakers exist and which speech belongs to whom. These are the same techniques used in published diarization research, not a proprietary black box.
Two details matter here. First, TranscribeX deliberately runs the pipeline at its higher-accuracy setting, scanning the recording with heavily overlapping analysis windows rather than the faster sparse pass. Second, you never have to tell it how many people were in the room: the speaker count is detected automatically, with no ceiling. A two-person interview and a twelve-person all-hands get the same treatment.
Once speakers are identified, each transcript segment is assigned the speaker whose speech overlaps it most, and consecutive segments from the same person merge into clean, colour-coded paragraph blocks—dialogue you can actually read.
Why it’s better than the usual approach
Private by physics, not by policy
Cloud diarization means your meetings—and your colleagues’ voiceprints—live on someone else’s servers. TranscribeX processes everything locally, so there is nothing to upload, retain, or breach.
Research-grade accuracy
The pyannote community-1 pipeline is a state-of-the-art open diarization system, and TranscribeX runs it at its most accurate configuration rather than the fastest one.
Fast, thanks to the Neural Engine
The models run as Core ML on Apple silicon’s Neural Engine. In the pipeline author’s benchmarks it processes audio at roughly 65× real time—about a minute for an hour-long meeting.
Unlimited speakers, zero setup
No speaker-count guessing, no per-seat limits, no configuration. Diarization can run automatically after every meeting recording.
On accuracy, the numbers are public. In benchmarks published by the pipeline’s maintainers, the exact offline pipeline TranscribeX uses averages a 10.6% diarization error rate on the standard AMI single-microphone meeting test set—matching the published results for pyannote community-1 itself—and detects the exact correct number of speakers in 12 of its 16 test meetings. They also measure the Core ML build at many times the throughput of the same pipeline running in PyTorch on a Mac’s CPU or GPU. You give up nothing by staying local; you gain speed and privacy.
There is a subtler accuracy advantage too: TranscribeX diarizes the finished recording as a whole rather than guessing speaker turns live in a stream. Batch diarization gets to look at the entire conversation before deciding who is who, which the pipeline’s own benchmarks show is meaningfully more accurate than streaming approaches.
Voice Profiles: never meet “Speaker 1” again
Separating voices is only half the job; the other half is knowing whose voices they are. That is what Voice Profiles do. After any diarization, select a speaker, and save them as a profile. TranscribeX stores that person’s voice fingerprint—plus a short audio snippet so you can audition it later—locally on your Mac.
From then on, every diarization automatically compares each detected speaker against all of your saved profiles and applies the matching name and colour. Your weekly stand-up transcribes itself with real names from the first second. A few refinements make this work well in practice:
- Save the same person more than once. Multiple samples—headset audio, conference-room audio, a phone line—improve matching accuracy, and TranscribeX groups them into one profile automatically.
- Choose how strict matching should be. Each profile has a match level—Strict, Normal, or Low—so you can demand near-certainty for sensitive recordings or cast a wider net for noisy ones.
- Organize profiles into groups. Keep your leadership team, your podcast guests, and your family as separate collections, each with their own names and colours.
- Everything stays local. Voice fingerprints are biometric data. TranscribeX keeps them in its local database on your Mac—they are never uploaded anywhere.
That last point deserves emphasis. A cloud service that recognizes your colleagues’ voices is, by definition, a cloud database of your colleagues’ biometric voiceprints. With TranscribeX, the convenience of automatic recognition doesn’t come with that trade-off.
From speaker labels to real answers
Diarization isn’t the destination; it’s what makes everything after it useful. Once speakers are named, TranscribeX puts those names to work across the whole app.
AI summaries that credit people. When segments carry real names, TranscribeX feeds the language model a speaker-attributed transcript, so meeting reports can say who made each decision and who owns each action item—not just that decisions were somehow made.
Chat with one voice at a time. Transcript chat can be scoped to a single speaker. Ask “what did the client actually commit to?” and get an answer drawn only from the client’s words.
Exports that keep the names. Speaker labels carry into every export format—TXT, Markdown, PDF, CSV, JSON, and SRT/VTT subtitles—and into the automatic post-meeting exports and emailed recaps. You can also group an export by speaker for interview-style documents.
An honest editing toolkit. No diarization system is perfect, so TranscribeX makes corrections cheap: rename a speaker inline, recolour them, merge two speakers the model split apart, select any set of segments and reassign them in bulk, or fix a whole paragraph with one click. A minute of cleanup beats an hour of “wait, who said this?”
On-device versus cloud diarization
Cloud notetakers can produce good speaker labels too. The difference is what your audio has to go through to get them.
| Consideration | TranscribeX | Typical cloud notetaker |
|---|---|---|
| Where diarization runs | On your Mac, via Core ML and the Neural Engine | On the provider’s servers |
| Audio leaves your device | No | Yes—uploaded for processing, often retained |
| Voice recognition data | Voice Profiles stored locally on your Mac | Voiceprints stored with the provider, where offered |
| Works offline | Yes, after a one-time model download | No |
| Speaker limits | Unlimited; count detected automatically | Varies by plan and provider |
| Ongoing cost | Included in the one-time $39 Pro license | Recurring subscription |
For plenty of teams a cloud workspace is a fine choice. But if your recordings include client calls, source interviews, medical or legal conversations, or simply meetings you’d rather not archive on a third-party server, “the diarization is good” isn’t enough—where it happens is the feature.
Where named speakers earn their keep
Teams in meetings
Action items with owners, decisions with names attached, and recaps that arrive by email before anyone forgets who promised what.
Journalists and researchers
Interviews separate cleanly into question and answer, and sensitive source audio never leaves the laptop.
Podcasters and video creators
Save each co-host and recurring guest as a Voice Profile once; every future episode transcribes as ready-to-publish, named dialogue.
Legal and compliance work
Attributable records of who said what, produced without sending privileged conversations to an outside service.
Using it in practice
Speaker diarization and Voice Profiles are part of TranscribeX Pro. For meeting recordings, diarization can run automatically the moment transcription finishes—detect, record, transcribe, diarize, summarize, export, email, all hands-free. For any other recording, open the Speakers panel and start diarization with one click; the models download once and run locally, offline, from then on.
It also scales past one file. Bulk imports have a diarization toggle for whole batches, and the tx command line and MCP server accept a diarize flag—so scripted pipelines and AI agents get speaker-labelled transcripts too.
Frequently asked questions
Does diarization work offline?
Yes. After a one-time download of the Core ML models, diarization, Voice Profile matching, and all storage run entirely on your Mac with no connection required.
How many speakers can it handle?
There is no fixed limit, and you don’t need to declare a count—the pipeline determines the number of speakers automatically.
What if it gets a speaker wrong?
Fix it in seconds: rename speakers inline, merge speakers that were split, or select segments and reassign them in bulk. Corrections flow through to summaries and exports.
Does it work with every transcription engine?
Local diarization runs after any local transcription engine, from Whisper to Parakeet. If you opt into a cloud engine that returns its own speaker labels, TranscribeX uses those instead of running the work twice.
When does diarization run?
After transcription completes—automatically for meeting recordings if you enable it, or on demand from the Speakers panel. Analyzing the full recording at once is part of why the results are accurate.
Is my voice data safe?
Voice Profiles—fingerprints and audio snippets alike—are stored only in TranscribeX’s local database on your Mac. They are never uploaded, and you can rename, regroup, or delete them at any time.
Put names on your next meeting.
Download TranscribeX, record one conversation, and watch it come back as dialogue—separated, named, and summarized, without leaving your Mac.
Editorial method: feature behaviour was verified against the TranscribeX application source on August 31, 2026. Accuracy and speed figures are from the diarization pipeline maintainer’s published benchmarks (AMI SDM and VoxConverse), measured on Apple silicon; results vary with hardware and audio quality.