Hinglish Transcription: How Accurate It Is and a 20-Minute Test

Hinglish transcription works, but expect mistakes: AssemblyAI rates Hindi at a 10–25% word error rate on its Universal-2 model, roughly one to two and a half wrong words in every ten, while ElevenLabs claims 5% or less for Hindi on its own tests. On sales calls the errors that matter fall on names, amounts, dates and phone numbers, so test any vendor on those with the 20-minute check below before you pay.

This applies whatever you are buying. A plain transcription tool, an AI call summary and a full conversation intelligence suite all start from the same transcript, and a wrong amount in the transcript becomes a wrong amount in the summary.

What "accuracy" means on a sales call

Vendors quote word error rate (WER): the share of words the software gets wrong, drops or adds, compared with what was said. A 10% WER means about one word in ten is off.

WER treats every word the same, and a sales call doesn't. A transcript of a 900-word call can be 97% right and still turn "₹1,48,500" into "₹1,40,500" or drop a digit from the customer's phone number. A missed "haan" costs nothing; a wrong digit loses the lead.

Two more cautions about the numbers you will see:

• They come from the vendors. ElevenLabs puts Hindi in Scribe's "Excellent Accuracy (≤ 5% Word Error Rate - WER)" band. AssemblyAI puts Hindi in Universal-2's "Good accuracy (>10% to ≤25% WER)" band. Neither describes a test on Indian sales calls.
• Benchmarks aren't phone calls. FLEURS, a common multilingual benchmark, is built on sentences from a machine-translation dataset, not on phone calls between a salesperson and a customer.

So count the words that matter instead: names, places, amounts, dates, times and phone numbers. The test below does exactly that.

Six things that break Hinglish transcripts

1. Phone audio. A mobile call carries far less audio detail than a studio recording. Sarvam, for example, says its speech-to-text handles "real call center audio: 8kHz, background noise, multiple speakers", 8 kHz being the narrow band phone networks use. Speakerphone, traffic and a weak signal make it worse.
2. Switching language mid-sentence. "EMI option hai kya, ya full payment karna padega?" is one sentence in two languages. Otter transcribes one language per conversation. Sarvam and Deepgram's Nova-3 multilingual mode are built for switching between Hindi and English.
3. Script. The same words can come out as Devanagari, Roman Hinglish or an English translation. Sarvam shows all three for one line: "मेरा फोन नंबर है", "mera phone number hai" and "My phone number is". Pick the version your team searches and types.
4. Numbers said in Hindi. "Dedh lakh", "saade teen hazaar", "unnis sau" and "adtees hazaar paanch sau" all have to become digits. These are the first words to check in any transcript.
5. Names and places. Surnames, localities like Wakad or Whitefield, and product names are rare words for a speech model. AssemblyAI and ElevenLabs both let developers supply a list of key terms to improve them.
6. Crosstalk. When both people talk at once, words get dropped or credited to the wrong person. Speaker labelling ("diarization") helps; Sarvam charges ₹45 an audio hour with it, against ₹30 without.

Which tools say they handle Hindi or Hinglish

These are the tools' own claims, from their official pages:

How we compared: languages, claims and prices from each provider's official page, checked in September 2026. Prices exclude 18% GST unless stated; USD prices are billed in USD.

AssemblyAI also supports Hindi: Universal-2 ($0.15 an audio hour, billed in USD) is the model in the 10–25% band quoted above, and the newer Universal-3.5 Pro ($0.21 an audio hour) covers Hindi with code switching built in. The API tools are raw engines. They need a developer, a way to collect call recordings and somewhere to show the text, so a small team without a developer meets them only inside a calling or call-analysis product.

The 20-minute test

Run this once per vendor you are considering. The first vendor takes about 20 minutes; each one after that about 10, because the recordings are reused.

1. Set up (2 minutes). Two colleagues, two phones. The "customer" sits somewhere noisy and uses speakerphone, as many real customers do.
2. Record the test call (5 minutes). Read the script below as a normal call, at normal speed, in your team's usual Hinglish. Don't slow down or over-pronounce.
3. Add two real calls (3 minutes). Pick two recorded calls from last week with plenty of numbers in them, from customers who heard the recording announcement.
4. Run all three recordings through the vendor (5 minutes). Use its free tier or trial credits, or ask the vendor to process them during a demo.
5. Score the 12 planted words (5 minutes). Use the answer key and score sheet below. For the real calls, score every amount, date and phone number.

Test call script

Illustrative script. The names, prices and phone number are made up.

TEST CALL (read as a normal call, at normal speed)

Agent:    Hello, main Imran bol raha hoon, packers and movers se. Aapne enquiry di thi?
Customer: Haan, main Bhagyashree Kulkarni, Wakad se. 2BHK shift karna hai Bengaluru, Whitefield.
Agent:    Pincode bata dijiye?
Customer: 560066. Date chaudah November rakhni hai.
Agent:    Theek hai. Saaman mein kya-kya hai?
Customer: Double-door fridge, ek 3-seater sofa, aur ek Honda Activa bhi jayegi.
Agent:    Toh total adtees hazaar paanch sau hoga, insurance alag. Advance paanch hazaar.
Customer: Thoda zyada hai yaar. Kal ek aur company ne chauntees hazaar bola tha.
Agent:    Ma'am, hum unpacking bhi karte hain. Chhattis hazaar final kar deta hoon.
Customer: Okay, done. Mera number note kijiye: 98201 45736.
Agent:    Subah saat baje team aa jayegi. Quotation aaj raat WhatsApp kar dunga.

Answer key and score sheet

Vendor: ______________     Recording: test call

 #  Planted word            Correct in transcript?
 1  Bhagyashree Kulkarni    Y / N
 2  Wakad                   Y / N
 3  Whitefield              Y / N
 4  560066                  Y / N
 5  14 November             Y / N
 6  Honda Activa            Y / N
 7  ₹38,500                 Y / N
 8  ₹5,000 (advance)        Y / N
 9  ₹34,000 (other quote)   Y / N
10  ₹36,000 (final)         Y / N
11  98201 45736             Y / N
12  7 am                    Y / N

Score: ____ / 12

Count a word as correct only if someone could act on it without the recording. "36 hazaar" or "36,000" counts; "3,600" does not. Small spelling slips in a name count if the name is still recognisable.

How we suggest reading the score:

• 11 or 12: fine for summaries and follow-ups. Still check amounts before sending a quote.
• 8 to 10: fine for the gist of a call. Never copy a number from it without the recording.
• 7 or fewer: not good enough for your calls. Try another vendor, or fix the audio first (headsets, no speakerphone) and test again.

When the transcript is borderline

If a vendor lands in the middle band, three habits make it workable:

• Read numbers back on the call, digit by digit. "Main confirm kar leta hoon: nine-eight-two-zero-one, four-five-seven-three-six." The customer corrects you, and the transcript gets a second, cleaner chance.
• Ask for addresses and email IDs in writing. A WhatsApp message beats any speech model at spelling "Kulkarni".
• Keep the recording one tap from the summary. When a number looks odd, the check should take ten seconds, not a search through a phone's file manager.

Once the transcript is good enough, the next choice is what to build on it: plain transcription and summaries, or conversation intelligence with rep scoring. For the ways to get calls transcribed in the first place, see phone call transcription options and costs.

For what a good summary should say, see the call summary guide. Voice AI that answers calls in Indian languages is a separate product, covered in our AI receptionist guide.

Put your own calls through TalkEasy

TalkEasy by GrowEasy records calls on your business number, transcribes each recording and writes a summary with the requirement, objections and next follow-up, usually within a couple of minutes. Before you rely on any tool for amounts and dates, ours included, run the 20-minute test with your own calls and score the result.

See how TalkEasy's AI call summaries work

Last reviewed 24 September 2026.