Every serious transcription tool now handles a clear recording of one person speaking at roughly the same accuracy, which makes published accuracy figures close to useless as a comparison. Business audio is not that. It is four people on a conference bridge, two of them talking over each other, one on a car speakerphone, using product names no model has seen. The tools separate on how they handle that, on whether sensitive content can be removed before storage, and on whether the recording behaviour is lawful in the states your team calls into. Those three questions decide the outcome, and none of them appears on a feature grid.
Overview: five things that decide a transcription tool choice
- Diarization is the hard part, not words. Attributing speech to the right speaker through overlap and crosstalk is where tools visibly differ, and a transcript with the wrong attribution is worse than no transcript.
- Vocabulary has to be taught. Product names, client names, drug names and internal acronyms are transcribed wrong until a custom vocabulary is supplied, and support for that varies.
- Consent law is a per state question. Several states require all parties to consent to recording, so a tool that starts recording silently creates legal exposure across state lines.
- Redaction has to happen before storage. Removing card numbers or health information after a transcript is written means it was stored unredacted first.
- Retention defaults are generous. Most platforms keep transcripts indefinitely by default, quietly building a searchable archive of every conversation the firm has had.
This is written for operations directors and IT leads at firms between 25 and 500 people rolling transcription across sales, support or clinical conversations. If the underlying phone platform is the open question, our explainer on how a VoIP phone system works covers that layer.
Why headline accuracy numbers do not predict business results
Word error rate on real audio diverges sharply from benchmark figures, because benchmarks use clean recordings of scripted speech and business calls are neither. A tool advertising ninety nine percent accuracy measured that on studio quality single speaker audio.
Overlapping speech and who said what
Diarization assigns each segment of audio to a speaker, and it degrades fast when two people talk at once, which is most of any real meeting. A transcript that attributes a client commitment to the wrong person is not a minor error, it is a record that says something untrue.
The counter position is that attribution errors are obvious to anyone who was in the room, so they self correct. That holds while a human reads the transcript soon afterward and fails completely once transcripts feed a summary nobody checks, which is the direction every deployment moves in. Test diarization on your own multi speaker recordings before committing.
Vocabulary the model has never encountered
Client names, product lines, internal acronyms and regional place names are transcribed phonetically until the tool is told about them. For a firm whose calls are full of proprietary terms, this is the single largest source of visible error.
Most platforms support a custom vocabulary or boost list, and the implementations differ in how many terms are allowed and whether they apply in real time or only in batch. This is a cheap fix with an outsized effect, and it is the first thing to configure rather than an optimisation to come back to.
Audio quality set by the phone platform, not the tool
Transcription quality is capped by what the capture layer delivers. A compressed narrowband codec, an aggressive noise suppressor cutting off word endings, or a headset that clips loud speech will limit every downstream tool equally.
Some argue this makes the tool choice largely irrelevant, which overstates it, and the underlying point is right: fixing the codec and the headsets often improves results more than switching vendors. That capture layer is part of the Teams phone system and platform work we usually do first.
Where the tool categories genuinely differ
Call recording consent and workflow shape split this market into three categories, and they suit different problems rather than being better or worse versions of each other.
Meeting assistants that join the call
Tools that join a meeting as a participant capture the room and produce notes without touching the phone platform, which makes them the fastest to deploy. They also appear in the participant list, which handles disclosure automatically.
The limitation is that they cover meetings and not telephone calls, and an external participant appearing in every client meeting is a conversation some firms have to have with their customers first. Bot presence is also a data path worth reviewing, because the assistant is a third party in the room.
Transcription inside the phone or contact platform
Where the phone system transcribes natively, the audio never leaves the platform already holding it, which removes a vendor, a data path and a contract review. For firms already consolidated on one communications platform this is usually the right answer.
What you give up is feature depth. Native transcription tends to lag standalone tools on diarization quality, language coverage and summarisation, and it is tied to the platform, so it becomes a switching cost later. Our piece on unifying communications under Teams covers the consolidation argument in full.
Speech APIs for building something specific
Where the requirement is structured output from calls at volume, such as extracting intent, flagging compliance phrases or routing follow ups, a speech API with summarisation and intent detection is the path. It is the only option that produces data rather than documents.
It needs engineering to build and maintain, which puts it out of reach for firms with no development capability. The trade is worth it when transcripts are an input to a process rather than a record for a person to read.
Getting the legal and retention side right before rollout
Transcript retention policy and consent handling are the two things that turn a productivity rollout into an exposure, and both need deciding before the first recording rather than after the first complaint.
All party consent states and calls that cross borders
Several states require every party to consent to a recording, and the rule that applies to an interstate call is not always the caller’s own state. A tool that begins recording without an audible notice is a problem for any team calling nationally.
The practical position is to configure an announcement on every recorded call regardless of jurisdiction, which is simpler than maintaining a per state matrix and removes the question. Some teams resist on the grounds that the announcement changes the tone of a sales call. It does, slightly, and it is cheaper than the alternative. The security side of the same platform is covered in our guidance on keeping a VoIP system secure.
Redaction that happens before the transcript is stored
A customer reading out a card number, or a patient describing a condition, is now in a text file that is searchable and backed up. Redaction applied after the fact means the unredacted version existed and probably still exists in a backup.
Tools differ on whether redaction runs in the processing pipeline before persistence or as a post process. For regulated content that distinction is the whole question, and it is worth asking the vendor directly rather than reading the feature name. Our cybersecurity compliance work treats this as a data handling decision rather than a product feature.
Retention that defaults to forever
Most platforms retain transcripts indefinitely unless configured otherwise, which builds an archive of every conversation the firm has had. That archive is discoverable, it is a breach target, and nobody decided to create it.
Set a retention period that matches why the transcripts exist. Notes to write up a meeting need weeks, not years. Compliance recordings have a defined period set by the obligation. The broader risk picture for communications platforms is covered in our piece on VoIP and unified communications risks, and the storage side sits with our cloud security practice.
Frequently Asked Questions
How accurate is AI transcription on real business calls?
Accuracy on clean single speaker audio is high and broadly similar across tools. On multi speaker calls with crosstalk, background noise and proprietary vocabulary it drops noticeably, and the drop varies by vendor, which is why testing on your own recordings matters more than published figures.
Do we need consent to record calls for transcription?
Several states require all parties to consent, and the rule applying to an interstate call is not always the caller’s own. Configuring an audible announcement on every recorded call is simpler than maintaining a per state matrix and removes the question entirely.
Can transcription tools remove sensitive information automatically?
Many offer redaction of patterns such as card numbers. The distinction that matters is whether redaction runs before the transcript is stored or afterward, because post processing means an unredacted copy existed and may persist in backups.
Is native phone system transcription good enough?
For firms already consolidated on one communications platform it usually is, and it removes a vendor, a data path and a contract review. Standalone tools generally lead on diarization quality, language coverage and summarisation depth.
How long should call transcripts be retained?
Match the period to the reason the transcripts exist. Meeting notes need weeks, while recordings kept for a compliance obligation have a period defined by that obligation. Leaving the default in place creates a searchable archive of every conversation nobody decided to build.
Who is behind this guidance
Mindcore has run communications platforms and compliance programmes for small and mid sized firms for more than two decades, across healthcare, legal, financial services and manufacturing clients. Most of our transcription work starts with the capture layer and the retention decision rather than with the tool, because those are what the managed IT and compliance sides of a deployment actually turn on.
Matt Rosenthal, our chief executive, focuses on how firms adopt convenience features without quietly creating records they did not intend to keep. His view on transcription is that the transcript is a business record from the moment it is written, and it should be treated like one from the start.
Test on your own audio before you choose
The evaluation that works takes an afternoon. Record a genuine multi speaker call with your usual equipment and run it through each shortlisted tool. Check speaker attribution through the parts where people talk over each other. Load your product and client names as custom vocabulary and see what changes. Ask each vendor whether redaction runs before storage. Set a retention period and an announcement policy before anybody goes live.
If you want help running that comparison against your own calls, book a free strategy call and we will look at the capture layer alongside the tooling.
