Skip down to the research record

Call transcript record: 4 fields before a deal review

Quota Forecast Editorial Team · Published · 13 min read

A call transcript record needs 4 fields before a deal review: call identity, speaker roster, speaker matching method, and redaction details.

Key takeaways

  • Copy the call ID and record the time zone and deal link.
  • Confirm attendees and their roles without guessing.
  • Record how speaker labels were matched to people.
  • Note redaction settings and unresolved gaps.

A revenue leader who reads transcripts during a deal review may be deciding whether a stated close date or a commit is backed by what the buyer actually said. If the record cannot answer those questions, the transcript is a document nobody can check. Use it alongside your commit evidence checks.

This post walks through 4 fields to complete on the record before the review starts. Two of the fields depend on how the transcript was produced. For those, the post quotes the public documentation of specific speech and transcript vendors and says which vendor's rule applies. Other vendors may work differently, so check your own tool's documentation before you assume the same behavior.

Each field answers one question a reviewer is likely to ask. Field 1 asks which call this is and where it sits in the deal. Field 2 asks who was in the conversation. Field 3 asks how the words were tied to those people. Field 4 asks what was changed before the text was saved. If the record answers all four, the reviewer can spend the meeting on the deal instead of on the paperwork behind it.

Field 1: the call identity and timing

The first field identifies the call. Write the date, the start time with its time zone, the deal or opportunity it belongs to, and the call ID that the recording system assigned. Fill these in before anything else, because a transcript without a date and a deal link cannot be matched to the stage history on the forecast.

Copy the call ID exactly as the recording system shows it. This lets a later reviewer search the recording system for the original and compare it with the text. Copying also reduces one kind of error: attaching a call to the wrong deal when two calls on the same day look alike.

Write the time zone next to the time. A start time of 9:00 means different moments when the people on the call are in different zones, and a reviewer comparing dates against a close date or a meeting calendar will read it differently if the zone is missing. Store the time in one zone across all records, and state that zone in the header of the record so nobody has to guess. Connect any changed close date to the close date change log.

Add a call type from a short list you define in advance, such as discovery, demo, pricing, legal review, or executive sponsor. The list is your own label, not something the recording tool sets for you. Use the same words on every record so a reviewer can filter a deal's calls by type without rereading each transcript. A short list is also easier to keep consistent than a long one, since every extra label is another choice that has to be made the same way each time. Agree on the list with the people who will read the records, and change it only when the change is written down for everyone.

Field 2: the speaker roster

The second field lists everyone who was on the call, with the company and role of each person. Build the list from the meeting invitation or the attendee list on the calendar event, not from memory after the call. Mark anyone who joined by phone only. If you know that someone joined late or left early, note that too, since a reviewer reading the transcript needs to know who could have said what.

A transcript often shows a speaker as a label rather than a name, so the roster is what connects a label to a person. Without it, the reviewer has text attributed to "Guest-2" and no way to tell whether that was the economic buyer, a technical evaluator, or a contractor. Field 3 explains how labels get matched to names. The roster is the list you match against, so it has to be complete before you start matching.

Separate internal and external participants. Put the seller's people in one group and the buyer's people in another, and note who owns the relationship on each side. A reviewer weighing a commit will want to see at a glance whether the person who signs the order was on the call. Put titles where that check is easy, not in a footnote.

Use the same role names your CRM uses for contacts, such as champion, economic buyer, or technical evaluator, if your team already has them. Role labels are only as reliable as the notes behind them, so a short line of evidence next to each label, such as a quote from the call or a message from the contact, will help the reviewer judge it.

Do not guess roles. If you do not know a title, write "unknown." A guessed title that later proves wrong damages trust in every other line on the record, while an "unknown" tells the reviewer exactly what to confirm before the review. If a buyer-side attendee was not named in the invitation, add that person once you know who they are, and record the date you made the edit.

Keep every roster entry in the same order and format: name, company, role, internal or external, and how the person joined. Consistency makes the record easy to scan, and it makes it possible to compare the attendees across several calls on the same deal. If the same person appears under two spellings, pick the one the CRM uses and note the other, so the reviewer does not count one buyer as two people.

The roster also needs a note about silent attendees. Someone listed on the invitation who never speaks in the transcript is not evidence of anything by itself. Write whether you know that person was present for the whole call, was present for part of it, or was only on the invitation. A reviewer who sees a senior name with no lines in the transcript should be able to tell whether that absence means something or only means the person listened.

Field 3: how each speaker was matched

The third field records how each speaker in the transcript was matched to a person. This matters because the method changes how much a label can be trusted, and the methods are not the same across tools.

Gong's documentation on identifying speakers says that Gong uses several methods to determine who was on a call and when each person spoke. Two of those methods are worth recording separately, because they work differently.

For telephony recordings, Gong's page says: "When Gong receives stereo recordings, we assume the two channels are the two speakers, and don't try to divide the call further." Under that approach, each channel is treated as one speaker. If a single phone line carried more than one person, the transcript will not split them apart on that basis alone, so your roster and your notes have to cover them.

For participants who join using their computer, the same Gong page says: "Participants identify themselves using their full name or a nickname." A nickname is not always the name the roster uses, so compare the participant's identification against the roster before you accept the match. Check that identification against the invitation.

The two quotes above describe separate identification methods. The stereo rule concerns channels, while the computer rule concerns participant identification. Record the applicable method without assuming the rules explain how a combined setup behaves. The record should show which speakers were matched by the channel, which by self-identification, and which were not matched at all.

Microsoft's quickstart for real-time diarization describes a different labeling scheme. It says: "Speakers are identified as Guest-1, Guest-2, and so on, depending on the number of speakers in the conversation." It also says: "The Speech service provides information about which speaker was speaking a particular part of transcribed speech." Under that scheme, the transcript gives you numbered guests, and the names have to come from somewhere else, such as your roster and your notes. The label tells you which part of the transcript belongs to which speaker. It does not tell you who that speaker is.

Whatever the tool, record three things for each speaker label:

  • the label as it appears in the transcript, such as Guest-1 or the vendor's own speaker tag;
  • the person it was matched to, using the names from the roster;
  • the method: channel, self-identification, the vendor's label, or a manual match by someone who listened to the call.

Manual matches deserve the most care. If you matched Guest-2 to a named person by listening to part of the call, record the timestamp of that part, the name of the person who made the match, and the date. A manual match is a judgment, and a reviewer can only weigh a judgment if the record shows that it was one.

Some speech will not be tied to any label you can confidently match. Record that as unmatched rather than filling the gap with a guess. A paragraph that a reviewer can see is unmatched is still useful, because it tells them where the record is thin. A paragraph that looks matched but was matched by guesswork is the kind of error that surfaces at the worst time.

If two labels appear to be the same person, or one label seems to cover two people, do not merge or split them silently. Write the conflict on the record, describe what you heard, and resolve it in the call notes. An open conflict is more useful to a reviewer than a clean-looking record that hides one.

Field 4: the redaction record

The fourth field records whether the transcript was redacted, which types of values were replaced, and how the replacements are numbered. Redaction changes the text a reviewer reads, so the record has to show what was changed. It also changes what a reviewer can search for. A reviewer looking for a customer's name in a transcript may not find it if that name was replaced with a tag, so the record should say that replacement happened and point to the tag that stands in for the name.

Deepgram's redaction documentation says you "may select the types of entities you wish to redact from over 50 supported entity types." Under Deepgram's redaction feature, then, what gets replaced depends on which entity types were selected when the transcript was produced. Write the selected types on the record. A reviewer who sees a tag and wonders what was there should be able to tell from the record whether that type was on the list at all.

Deepgram's example shows how a replacement appears in the text. With redact=ssn, a spoken social security number is replaced with a tag. The example in the documentation reads: "My social security number is [SSN_1] and his is [SSN_2]." The number inside the tag is an index, so the reader can tell that two different values were replaced, not one value repeated.

Deepgram also says that indexes persist for the whole stream: "a value tagged [NAME_1] early in the stream keeps that tag in later segments, and new values continue the count." Use that rule to check the record. If [NAME_1] appears in one place and a different tag appears later for what sounds like the same person, write that down as a question rather than resolving it yourself.

Deepgram's documentation also says that numerals matching an entity type in the numbers group are tagged with the specific entity, such as [CREDIT_CARD_1] or [SSN_1]. A card number read aloud on a call may therefore show up as a tag instead of digits, and a reviewer cannot check the figure from the transcript alone. Note in the record that digits were replaced, so the reviewer knows the figure has to be confirmed some other way, such as by asking the person who was on the call, rather than assuming the text is complete.

If you run Deepgram yourself, the release matters. Deepgram's redaction page says that self-hosted number redaction "requires release 260430 or later; Bulgarian, Chinese, and Cantonese require 260611 or later, Malay 260714, Korean 260728, and Japanese 261001 (language=multi requires 260611 or later)." Record the release your self-hosted service runs and the language of each call, so a reviewer can check the record against those requirements. This post does not cover the requirements of any other vendor, so look those up in that vendor's own documentation.

Finally, do not read a redacted transcript as proof that nothing sensitive remains. Spot-check the part of the call where the buyer discussed contract terms or payment details. If you find a value that should have been replaced and was not, record it as a known gap, and tell the person who owns the redaction settings so they can review it.

If your system keeps a link to the original audio, add that link to the record as well. A reviewer who doubts a line in the redacted transcript can then check the recording, and the record should say which of the two the reviewer is reading. Keeping both in view is the simplest way to avoid arguing about a sentence that the text itself cannot settle.

Before the review, check all 4 fields

Before the review, check that all 4 fields are filled in, and that a reader who never heard the call can answer three questions from the record alone: which call this was, who was on it and which speaker matches which person, and what was replaced in the text. If an answer depends on memory, write it down now, while the call is still fresh for whoever attended it.

A complete record for one call should contain:

  • Field 1: the call ID copied from the recording system, the date and start time with the time zone, the deal ID, and a call type from the agreed list.
  • Field 2: the roster of named attendees, each with company, role, internal or external status, and how they joined.
  • Field 3: one line for each speaker label, with the method used and the timestamp and author of any manual match.
  • Field 4: whether redaction was on, the entity types selected, the tags that appear in the text, and the release number if the service is self-hosted.

Use "unknown" for any blank that cannot be filled before the review, and name the person who will fill it in. A blank looks like an oversight. An "unknown" with an owner gives the reviewer a clear next step and shows that the gap was seen.

Keep the record and the transcript in sync. If someone corrects a speaker label, a roster entry, or a redaction setting after the review, log the change with the date and the name of the person who made it. A forecast record that changes without a trail is harder to defend later than one that shows its edits, and the point of attaching a transcript is to make the forecast easier to check, not harder.

Illustrative example: checking a speaker match

The record already identifies the call and its deal. Instead of assigning the label by guesswork, the reviewer marks it unmatched and names the attendee who will check the audio. After confirmation, the edit records the matching method, timestamp, author and date.

Today, open the record for your next deal review and check its speaker matches against the roster. Assign an owner to any unresolved match before the meeting.

Questions before the review

Does a speaker label identify a person?

A speaker label alone does not establish a person's identity. Keep the roster and matching method beside the transcript so the reviewer can check the attribution.

Is a redacted transcript proof that sensitive details are gone?

A redacted transcript is not proof that nothing sensitive remains. Use the redaction record and the spot-check described above to identify gaps before relying on the text.

Sources