Skip across the chapter
Quota Playbook

Wrong-language call transcripts: 4 checks to run

Quota Playbook Editorial Team · Published · 12 min read

Wrong-language transcripts need four checks: confirm the spoken language, check for mixed languages, reprocess in Gong, then look for patterns.

You open a recorded call, expecting a clean transcript for coaching, and instead get a wall of text in a language nobody on the call spoke. Or you get a transcript that is half right, with the prospect's sentences rendered as nonsense. Automated transcription makes this happen, and when it does, the coaching you planned for that call stalls.

Key takeaways

  • Listen to the call first and write down the language each speaker uses.
  • Ask whether the call mixed languages, then choose the language the review needs.
  • Reprocess within 90 days of the recording date, and expect the call to be unavailable for up to an hour.
  • Log each incident so repeated failures show up as a pattern.

On a small sales team, this hurts more than it would at a large company. There is usually no operations person whose job is to fix recording problems. The manager who runs weekly call reviews is also the person who notices the problem, diagnoses it, and decides what to do. That is manageable if you have a simple routine for it.

The good news is that this problem is usually quick to diagnose and often quick to fix. The harder part is that the fix depends on what actually went wrong. A call can be transcribed in the wrong language because the system guessed incorrectly, because the call switched languages partway through, because the recording was set up without a language, or because the system consistently mislabels a certain kind of call. Each cause has a different remedy, and applying the wrong one wastes time.

Below are four checks to run, in order, before you decide a recording is lost. They are written for front-line managers of small B2B software sales teams, where a handful of reps and a few recurring call types make patterns easy to spot once you look for them.

Check 1: Confirm which language was actually spoken

Start by listening. Before you touch any setting, play the first few minutes of the call and the parts where the prospect speaks. Write down the language you hear for each speaker. This sounds obvious, but it is the step most often skipped, and it determines everything that follows. Do not rely on the transcript to tell you what was said, because the transcript is the thing you are testing.

As you listen, pay attention to audio quality too. Background noise, a poor headset, a speakerphone in a busy room, or a call with frequent dropouts can all degrade a transcript, and they can make a correct language guess look like a wrong one. If the audio is hard for you to follow, the system may struggle as well. Note this in your log, because it changes how much weight you give the transcript later.

There are three outcomes worth distinguishing:

  • The call was entirely in one language, and the transcript is in a different one. This is a detection problem.
  • The call was in one language, but the transcript is garbled or partly nonsense. This may be a dialect or accent mismatch rather than a wrong-language problem.
  • The call moved between languages. This is a multilingual call, and a single transcript usually cannot be fully right in both languages unless the tool handles mixing well.

Gong's help page describes the default behavior in one sentence: "Gong automatically detects the language spoken on a call and generates the transcript in that language." ( Gong help: Transcribe a call in the correct language) That sentence explains the first outcome. The system makes a judgment about which language to use, and it can get that judgment wrong. Knowing this shifts your mindset. The transcript is the output of a guess, not an objective fact about the call. When you see the wrong language, the useful question is not "what broke in the recording?" but "did the guess fail, and can I redo it?"

Write your finding in one line, such as "Entirely Spanish, transcribed as Portuguese" or "Rep in English, prospect switching between English and French." You will need this for checks three and four. Keep the line factual and short. You are recording what you heard, not what you think the rep should have said.

Check 2: Ask whether the call contained more than one language

Bilingual and multilingual calls are common on small B2B teams that sell across regions or serve customers who switch languages mid-conversation. A rep might open in one language and move into another when the prospect is more comfortable, or a customer's technical team might answer in their working language while the buyer speaks another. A single transcript cannot always capture both cleanly.

Gong's guidance covers this case directly: "If the detection was wrong, or if the call included multiple languages and you want the transcript in a different one than the one that was generated, you can reprocess the call from the call page." ( Gong help) This tells you two things. First, the tool expects that a call may need a specific language chosen after the first pass, even when that pass looked reasonable. Second, the choice is yours. You decide which language the transcript should serve.

So ask a practical question: what is this transcript for? If the goal is to coach the rep on discovery questions, the language the rep used for discovery should be the target, even if the prospect answered in another language. If the goal is to review what the prospect objected to, the target may be the prospect's language. There is no universal right answer. Pick the language that matches the purpose of the review, and apply the same rule across your team so that comparisons between calls remain meaningful. A manager who reviews one call in Spanish and the next in English, with no reason recorded, will struggle to spot real trends.

Dialect and regional variation deserve a separate note. Speech vendors often handle regional variants differently from languages. AssemblyAI documents that its Universal-3.5 Pro model "goes beyond standard language support with deep understanding of regional dialects and local variants." It also says the model "automatically recognizes regional speech patterns when using the base language code," such as fr for French or es for Spanish. ( AssemblyAI supported languages) If your tool is built on a vendor like this, a call that sounds like the right language but produces the wrong words may be a regional-variant issue, which is a different problem from a wrong-language detection. Read the vendor's language documentation before assuming every bad transcript has the same cause.

The same vendor page shows how language detection can be configured in the developer API. It explains how to "automatically fall back to Universal-2 for anything outside this set" by setting the model list and turning on language detection. ( AssemblyAI supported languages) You may never touch these settings yourself, but if your tool is built on AssemblyAI, they may explain why a transcript comes from a different model than you expected. If your team's recording tool exposes model or language options, ask your admin which ones are active.

Accuracy also varies from language to language. AssemblyAI's page says it "groups Universal-2 languages by transcription accuracy, measured by Word Error Rate (WER)." ( AssemblyAI supported languages) The practical lesson is that the same tool can perform noticeably differently depending on the language. If your team sells in several languages, a low-quality transcript may reflect the model's performance in that language rather than anything about the call. Treat coaching conclusions drawn from low-accuracy transcripts with more caution, and lean more on your own listening for those calls.

Check 3: Reprocess the call and verify what changes

If listening confirmed that the transcript is in the wrong language, reprocess the call. In Gong, reprocessing is started from the call page, and the official guidance describes several things to know before and after you do it.

First, check the window. Gong says reprocessing works "as long as you are within 90 days of the call recording date." ( Gong help) If a call is older than that, reprocessing may not be an option, and you will need to rely on your notes or the original recording. Build the habit of checking language issues soon after a call, not weeks later, when the window may have closed. A good routine is to scan new recordings within a day or two as part of your normal review schedule.

Second, expect a new transcript and re-run analysis. Gong states that "when you reprocess a call, Gong generates a new transcript in the language you select and reapplies its analysis." When the work finishes, "the transcript is updated in the selected language and Gong reapplies its analysis, including topics, trackers, and AI summaries." ( Gong help) This matters because any summaries or tracker results you read earlier were built on the wrong transcript. Once reprocessing finishes, go back and read the summary and the tracker hits again. Do not assume the old notes still apply. If you shared a summary with a rep, a leader, or a deal review, tell those people it has changed, so nobody is working from the outdated version.

Third, plan around downtime. Gong warns that "the call is unavailable during processing, which can take up to an hour." ( Gong help) For a manager, this is a scheduling issue. Do not reprocess a call five minutes before a coaching session where the rep expects to pull it up. Start the reprocess when you are not planning to use the call right away, or tell the rep the call will be unavailable for a while so they are not caught off guard. For the session itself, see Sales One-on-One Agenda for Coaching Software Sellers.

Fourth, verify the result. Scroll through the new transcript for the first few minutes and check that it reads correctly in the language you selected. Spot-check a section in the middle as well, since the opening can be right while later segments are not. If the transcript is in the language you chose but still unusable, the problem is probably not detection alone, and you should move to Check 2, which covers dialect and accuracy causes.

If calls are recorded through Microsoft Teams, Microsoft's guidance on recording meetings says to "specify your spoken language and what you want to record" when you start recording. ( Microsoft Support) Setting the language at the start avoids relying on automatic detection. Ask your team to treat language as part of recording setup rather than something to sort out afterward. Where the setting is available, a few seconds of attention at the start of a call can save a manual reprocess later and a stalled coaching session.

Check 4: Decide whether this is a one-off or a pattern

A single wrong-language transcript is an annoyance. A recurring one is a process problem. The fourth check is about telling the difference and then acting on it.

Start a simple log. You do not need a complicated system. A spreadsheet with one row for each incident is enough. For a related note format, see Sales call review. Capture the date, the rep, the account or call type, the language the system generated, the language actually spoken, whether the call contained more than one language, audio quality notes, and what you did to fix it. After a few weeks, patterns become visible quickly. You may find that one rep's calls are consistently misdetected, that calls with a particular customer region fail more often, or that multilingual calls are the main source of trouble rather than detection itself.

Each pattern suggests a different response:

  • If one rep's calls keep failing, check whether that rep switches languages more than others, has a particular accent, or records under noisier conditions. Coach on recording setup, such as stating the working language early in the call or keeping the audio environment consistent, rather than on the rep's speech. Frame this as a setup issue, because setup is the first thing to check. Reps who feel blamed for how they talk tend to stop taking the coaching seriously.
  • If calls in one language consistently come back with poor quality, the issue may be the model's accuracy in that language, which AssemblyAI documents in its language tables. Treat those transcripts as less reliable for coaching, and rely more on your own listening and notes for those calls.
  • If multilingual calls are the main problem, decide as a team which language each call type should be reviewed in, and write that rule down so every reviewer applies it the same way.

Escalate when the problem is consistent. Gong's guidance is direct on this point: "If Gong consistently identifies your calls in the wrong language, contact Gong support." ( Gong help) Do not wait for a pile of incidents before raising it. Once you can show a pattern, send support a few concrete examples from your log, including the call dates, the language detected, and the language actually spoken. The more specific your report, the easier it is for the vendor to investigate, and the less time you spend going back and forth.

Keep privacy and consent in mind throughout. Reprocessing and sharing a transcript can spread a customer conversation further than it was originally meant to go. Confirm that your team's policies cover who can access the recording and its transcript, and limit sharing to people who need it for coaching or deal review. When you forward a summary, make sure it reflects the current, reprocessed version, and avoid pasting full transcripts into channels where they do not belong.

Illustrative example

A manager reviews 10 calls in a week, and 3 come back in the wrong language, all from one rep. Logging them shows a pattern, so the manager checks that rep's recording setup first, then reprocesses each call within 90 days.

Wrap-up

A wrong-language transcript is rarely a dead end. Confirm what was spoken, check whether the call mixed languages, reprocess within the window and verify the new output, and then look for patterns before assuming the problem is isolated. Done in that order, these four checks turn a confusing transcript into a clear decision: fix this call, change how recordings are set up, or escalate to the vendor. Your coaching stays on track, and your team keeps a reliable record of what was actually said.

Next step for today

Today, open your most recent call and write down the language each speaker used. Compare your note with the transcript. If they differ, decide whether to reprocess the call before your next coaching session.

Frequently asked questions

How long do I have to reprocess a call?

Reprocessing works only within 90 days of the call recording date, so scan new recordings within a day or two.

Will the call be usable while it reprocesses?

No. Expect the call to be unavailable for up to an hour, so start the reprocess when you do not need the call right away.

When should I contact Gong support?

Contact Gong support once a pattern shows up in your log, not after a single bad call.

Sources

  1. Product pricing and deal fields: 9 checks for coaching
  2. Sales text rules: 4 checks before a follow-up
  3. Service-dog questions: the 2 you may ask at the office
  4. Rep call score averages: 3 values to compare
  5. Transcript word boosting

More plays