Transcribing Bangla-English sales calls: what makes them hard and what we do about it

Listen to a sales call in Dhaka and count the languages. A rep opens in Bangla, quotes the price in English, names the package in English, and the customer answers in a mix of both. “Monthly package ta koto?” is one sentence and two languages.

That is normal speech for a lot of sales teams, and it is hard to transcribe. When we built AI Call Analysis, getting these calls right mattered more than anything else, because every score and recommendation depends on the transcript.

Why mixed-language calls are hard

Two languages in one sentence. Transcription that is set to a single language struggles here. Set it to Bangla and the English words come out garbled. Pick English and you lose most of the call.

Words no dictionary knows. Your brand, your package names and your product terms are often invented words. A transcriber has never heard them and will happily turn them into something that sounds close.

Two people on a phone line. A recording is one audio track with two voices. If you cannot tell who said what, “I want to start this month” could be the lead or the rep.

Calls that are not really calls. Plenty of recordings are a ring, a hang-up, a voicemail, or ten seconds of “hello? hello?”. Scoring those as sales conversations produces nonsense.

Readers who prefer different languages. A manager may want English summaries; the floor may read Bangla faster.

What we do about it

Bangla and English hints together

Every call is sent to speech to text with both Bangla and English as language hints, so a single call can switch between them without being forced into one. The scoring step is also told plainly that calls are usually in Bangla, often mixed with English, and that the transcript comes from automatic recognition, so it should expect some errors and imperfect speaker labels.

Words the transcriber should know

In Settings → Call Analysis there is a field called Words the transcriber should know. Put your brand, package and feature names there, comma separated. They go to speech to text as vocabulary hints on every call, which gives the transcriber a fair chance at the words that matter most in a sales conversation.

Separating the speakers

Speaker separation (diarization) is on for every call, so the transcript comes back as turns: speaker 1, speaker 2. Then the scoring model works out which speaker is your agent from context, since the agent is the one introducing the company and asking questions. On the call panel the transcript shows Agent and Lead instead of numbered speakers, and you can copy it as plain text with timestamps.

Skipping calls that are not conversations

Three filters keep junk out:

• Too short. Automatic analysis skips calls shorter than Skip calls shorter than in settings (30 seconds by default). You can still click Analyze call on any recording by hand.
• Near silent. If a transcript has fewer than 12 words, the call is marked as no real conversation, and the scoring model is never called.
• Not a conversation. If the model sees voicemail, a call that did not connect, or almost no real exchange, it marks it as not a conversation. Those calls never count toward a lead’s status.

Summaries in English or Bangla, quotes as spoken

Pick English or Bangla under Language in Settings → Call Analysis. It applies to everything the AI writes: the summary, the note on each signal, the next step, the questions to ask on the next call, the coaching tip, and the status labels on the pipeline board, the deals list, the contact page and the call panel.

What it does not touch is what people said. Transcripts stay as spoken, and the quotes attached to each signal are copied from the transcript without translation. A manager reading an English summary still sees the customer’s own words underneath, in whatever mix of languages they used.

Changing your mind later

Settings change. You might start in English and switch to Bangla, tighten your product description, or adjust the thresholds. New calls pick up the change right away. Calls already analysed keep their old language and scores until you re-score them.

Re-score existing analyses in Settings → Call Analysis runs the scoring step again from the saved transcripts. Nothing is transcribed again. First it does a dry run that shows how many calls are in the window and how many will be scored. Then it scores them with your saved language and settings and updates each contact. Re-scoring does not count against your daily limit or your monthly analysed calls. Save your settings first, since it uses the saved language, not what is on screen.

Recordings are deleted after 45 days, but by default the transcript is kept for longer, so re-scoring still works after the audio is gone. Once a transcript is purged at the end of its retention period, that call keeps its last result.

Where to start

Set up recording first with the call recording setup guide: MicroSIP on Windows plus the NerdCRM recorder agent. Then turn on call analysis in shadow mode, fill in what you sell and your key terms, and check a few transcripts from calls you know. Pricing for the add-on is on the pricing page.