Call Transcription for Indian Languages
Jaseem Abbas
- Published

Call transcription for Indian languages is the job of turning a phone call in Hindi, Tamil, Bengali or a Hindi-English mix into written text you can search, check and act on. A transcript in English is one thing. A transcript of a Tamil renewal call that mixes in English words, on a noisy mobile line, is another. This guide is for operations and quality teams that need to buy, test or run it, and it covers what changes, which script to store, and how to check a transcript on your own calls.
Here is a call to picture. A Chennai insurance renewal desk phones a policyholder, who says, "Premium amount பத்தாயிரம், next week கட்டுவேன்" (premium amount ten thousand, I will pay next week). A useful transcript keeps the Tamil words in Tamil script and the English words in English. It writes the amount, 10,000, as digits in its own field. It also flags "next week" as a promise with no date, so someone asks for one. A weak transcript turns half the line into guesses and loses the amount. The desk, the policyholder and the numbers are an example we made up. They are not customer data.
This post is about the transcript itself, across languages. For the detail of Hindi accuracy on phone lines, see our guide to Hindi speech to text on calls.
What is call transcription?
Call transcription turns the audio of a phone call into written text, either live during the call or afterwards from the recording. Twilio, a phone platform, defines it the same way: it "converts the audio from a voice or video call into written text" (Twilio on call transcription, read 15 November 2026). The software underneath is automatic speech recognition, and the output is the same whether the call is a sale, a reminder or a complaint.
Three words are often mixed up:
- Transcription writes down what was said, in the language it was said.
- Translation turns it into another language. It is a separate step.
- Transliteration shows the same words in other letters, such as Hindi in Roman letters.
Keep the three apart in your tools and your reports. A transcript that quietly mixes them cannot be trusted.
Why is call transcription harder in Indian languages?
Indian calls mix languages, accents and scripts on narrowband phone audio, and most speech models saw far less of that than of clean, studio-style English. Three things stack up on a real call.
- The line. A phone call carries a thin slice of the sound a microphone would catch. Our Hindi speech to text guide explains the phone band and what research says about it.
- The mix. Callers move between their language and English inside one sentence, and "code-mixing" of that kind is normal on Indian floors. See our glossary entry on code-switching.
- The spread. The 8th schedule of India's Constitution lists 22 languages, each with its own accents by district.
Research gives a sense of scale. The IndicVoices project collected 7,348 hours of speech from 16,237 speakers in 145 Indian districts across 22 languages, and 1,639 hours of it was transcribed. Of that speech, 9% was read aloud, 74% was extempore and 17% was conversational. The authors describe IndicASR as the first speech model to support all 22 languages in the 8th schedule of the Constitution (Javed et al., IndicVoices, read 15 November 2026). The point for a buyer: real, natural speech is the hard part, and it is a small share of what exists.
How should Hinglish and other mixed speech be written down?
Write it as it was spoken: Hindi words in Devanagari, English words in English, with no silent translation. That keeps the transcript true to the call, and it keeps names, product words and amounts searchable.
Here is one line, written three ways.
Style: As spoken
- Text
- हाँ जी, कल EMI कट जाएगा, but account में balance कम है
- Problem
- None. It matches the call
Style: All Roman letters
- Text
- haan ji, kal EMI kat jayega, but account mein balance kam hai
- Problem
- Spellings drift, so search misses
Style: Forced into English
- Text
- Yes, the EMI will be deducted tomorrow, but the balance is low
- Problem
- A translation, and it hides the original
The third style looks tidy and is the most dangerous, because nobody can tell what the caller really said. If a reader needs English, add it as a separate, labelled translation field. Our guide to Hinglish voice agents goes deeper on mixed speech on calls.
Which script should a transcript use?
Use the script of the language spoken, and add a transliteration or translation as a separate field only if a reader needs it. This is our advice. It means the transcript of a Gujarati call is in Gujarati script, and a search for a Gujarati word finds it.
Language: Hindi
- Script
- Devanagari
Language: Marathi
- Script
- Devanagari
Language: Bengali
- Script
- Bengali
Language: Gujarati
- Script
- Gujarati
Language: Punjabi
- Script
- Gurmukhi
Language: Kannada
- Script
- Kannada
Language: Malayalam
- Script
- Malayalam
Language: Odia
- Script
- Odia
Language: Tamil
- Script
- Tamil
Language: Telugu
- Script
- Telugu
Language: English (India)
- Script
- Latin
Two more rules help. Write numbers as digits wherever the business acts on them, such as an amount or a date. And keep a word list of brand names and local place names, so the same name is spelt one way every time.
Should you record one channel or two?
Two channels, one per speaker, give cleaner transcripts and correct speaker labels. Twilio's documentation says a single-channel recording stores one mixed track, which "can increase the eventual number of transcription errors, especially if participants are speaking at the same time". Dual-channel recording "performs better at eliminating cross-talk" and "prevents most (or all) misattribution errors" (Twilio, read 15 November 2026).
On a collections or sales call, where both sides often talk at once, that matters. If a customer says "no" over the agent's offer, one mixed track can give the wrong speaker the word. Ask your carrier or recording tool whether it can record the caller and the agent separately. Our guide to call recording software in India covers how to choose one.
How accurate is call transcription in Indian languages?
It depends on the language, the topic and the line, so the only number that matters is the one you measure on your own calls. The Vistaar study built 59 benchmarks across Indian language and domain pairs, and its authors argue that diverse benchmarks are needed because one test does not show how a model does elsewhere (Bhogale et al., Vistaar, read 15 November 2026).
Treat any single percentage with care. A figure measured on read-aloud news says little about an angry caller on a train. Word error rate (WER) is the usual score, the share of words that are wrong, and our Hindi guide shows how to measure WER and CER on your own calls. For the full benchmarking method, read benchmarking speech-to-text on Indian calls.
How do you test a transcription tool on your own calls?
Run about 30 real calls per language through it, and check the fields you act on: names, amounts, dates and addresses. The number 30 is our rule of thumb, not a standard. It is enough to show a pattern without a month of work.
- Pick the calls. Take 30 real calls per language you serve, including your noisiest hours and a few from weak networks.
- Choose the fields. Pick the ones the business acts on: amount, date, PIN code, name, and a clear yes or no.
- Write the truth. Have a person who speaks the language write what the key fields should say.
- Compare line by line. Read the tool's transcript beside the truth for each call.
- Score the fields first. Count how many calls got the amount, the date and the name right. Overall accuracy comes last.
- Check the script. Confirm Hindi came out in Devanagari and English words in English, as you decided above.
- Repeat after a change. Run it again when you change the tool, the carrier or the recording setting.
A transcript with 95% of words right can still get the amount wrong. That is why fields come before a single percentage.
What can a call center do with transcripts?
A transcript lets a team search calls, audit quality, write call summaries and fill CRM fields without replaying audio. That is the real value, and it is why the first fix is the transcript itself.
- Search. Find every call where a customer said "cancel" or named a rival, in any language you serve.
- Audit. Score a sample of calls against a checklist. Our guide to call audit in BPO has one.
- Analyse. Spot patterns across thousands of calls in speech analytics.
- Summarise. Turn a call into a short note for the next person, as in our guide to AI call summaries.
- Fill fields. Write the promised date and amount into the CRM, instead of asking an agent to type them.
Searches for "call recording analysis" mean the same job: read the calls without listening to every one.
Do you need consent to transcribe calls in India?
Rules on recording and transcribing calls vary, so check them for your own case. Twilio says that "call transcription legality differs by jurisdiction" and that some places "require some or all parties in a conversation to provide informed consent". This is not legal advice. Our guide to whether call recording is legal in India and the DPDP rules for call centers cover what applies to a floor like yours.
How Bananaflow fits
Bananaflow is an AI voice agent platform. For the calls its agents take, these parts touch the transcript:
- English (India) and 10 Indian languages: Hindi, Bengali, Gujarati, Kannada, Malayalam, Marathi, Odia, Punjabi, Tamil and Telugu, each heard and spoken in its own script. See Languages.
- Mixed speech written as spoken. Hindi with English words arrives as the caller said it. See mixed speech.
- The agent follows the caller's language and switches mid-call once the language is clear.
- A speech-to-text word list for brand names and jargon, so the agent hears them.
- A transcript and a recording after each call. Call QA can score a finished call and return tags, sentiment and a summary.
- Call data is not used to train or fine-tune models.
Test on your own calls, as above, with the lines and accents your callers bring.
What should you do next?
Pick one language and one call type this week. Pull 30 real calls, write down the fields you act on, and score a transcript against them. Decide the script rule before you compare tools, so every tool is judged the same way. A transcript is only as useful as the fields you can trust in it.
FAQ
Frequently asked questions
Can calls be transcribed in Hindi?
Yes. Hindi call transcription writes down spoken Hindi, best in Devanagari with English words kept in English, as the caller said them. Phone lines and Hindi-English mixing make it harder than a clean recording, so test on your own calls and score the names, amounts and dates first.
Which Indian languages can call transcription handle?
It depends on the tool, so ask for the exact list and test each language on your own calls. Research such as IndicVoices covers 22 Indian languages. Bananaflow agents work in English (India) and 10 Indian languages: Hindi, Bengali, Gujarati, Kannada, Malayalam, Marathi, Odia, Punjabi, Tamil and Telugu.
What is the difference between transcription and translation?
Transcription writes down what was said, in the language it was said. Translation turns it into another language. Keep them as two separate fields, so the original words stay on record and a reader can still get the English meaning.
Is call transcription accurate on phone lines?
It is usually less accurate than on a clean recording, and how much less depends on the language, the topic and the line. No single published number covers your calls. Run about 30 real calls per language and score the fields you act on, such as amount, date and name, before you trust a figure.
Is it legal to transcribe calls in India?
Rules vary by case, and this is not legal advice. Twilio notes that call transcription legality differs by jurisdiction and that some places require some or all parties to give informed consent. Read our guides on call recording law and the DPDP rules, then confirm with your own adviser.
What is dual-channel recording?
Dual-channel recording keeps each speaker on their own audio channel, so the caller and the agent are separate. Twilio's documentation says it performs better at eliminating cross-talk and prevents most misattribution errors, compared with one mixed track.




