Hindi-English Code-Switching on 8 kHz Helpline Calls

Executive Summary

A caller to a state helpline rarely speaks one clean language. A Hindi sentence carries English nouns for documents, departments and schemes, and the audio reaches the speech engine as an 8 kHz telephone signal. Both facts are routine, and published work shows each one making recognition harder in ways a language-by-language demo on a quiet line does not reveal.

This guide is for the state IT department, NIC unit or helpline cell that runs an IVR menu and a team of agents today. The caller picks a language, an agent listens and types the complaint into the portal. The change proposed here is to build one extra test set before a voice agent takes any of those calls: recorded, code-switched, 8 kHz speech from the department's own line, scored in four separate ways.

Executive Callout The published evidence is thinner than vendor demos suggest. We found no study that scores a current commercial or Bhashini engine on spontaneous Hindi-English telephone speech. The Hindi-English telephone corpora we found are recorded at 8 kHz, but they are read sentences scored on older systems. What the literature does establish is four places to look: the audio channel, the output script, the way errors are counted, and numbers and names. AiSewak is an implementer that builds on services like these. It is not the Bhashini portal and claims no endorsement or partnership from it.

Introduction: The Caller Who Says "Application Status" in Hindi

Picture a call to a district grievance line. The caller says, in Hindi, that the status of an online application for a ration card has not changed, and gives a complaint number. In everyday speech several of the nouns in that sentence would be English words, and the number may be spoken in either language. This example is illustrative, not drawn from a call recording.

An agent handles such a call without noticing the switch. A speech engine does notice it, because the engine was configured for a language and the sentence belongs to two.

Researchers have studied this for years. A 2018 corpus paper from IIT Guwahati calls switching inside a sentence "a common phenomenon" in bilingual communities and says switching between Hindi and English is very common in India. The same group's later work lists the engineering consequences. This guide uses that literature, plus AI4Bharat's benchmarks, to turn the problem into a test a procurement team can run.

What the Research Corpora Show, and What They Cannot

Every study below recorded or scored Hindi-English or Hindi telephone speech. None used a current commercial engine or a Bhashini model, so none ranks vendors.

Source	Audio	Result	What it tells a buyer
Sreeram, Dhawan and Sinha, Hindi-English Code-Switching Speech Corpus (arXiv 1810.00662, 2018)	Speakers called a toll-free number from mobile phones in homes and offices; 8 kHz; 7,005 utterances from 71 speakers	A DNN-based system scored 25.40% word error rate	A Hinglish corpus recorded the way a helpline hears it: over a phone call
Dhawan et al., HingCoS experiments (arXiv 1907.08293, 2019)	8 kHz, 16-bit; 9,251 utterances, about 25 hours, 101 speakers	Compared ways of shrinking the output label set	The same IIT Guwahati group again chose 8 kHz for this task
MUCS 2021 challenge (arXiv 2104.00235)	About 600 hours across seven Indian languages plus Hindi-English and Bengali-English pairs	Baseline word error rate of 30.73% on the multilingual task and 32.45% on the code-switching task	The two baselines landed within two points of each other; the paper does not isolate code-switching as a cause, but it did warrant its own challenge track
Vistaar (arXiv 2305.15386, 2023)	Hindi telephone-quality speech from the NGO Gramvaani	Hindi error 7.6 to 26.8 for AI4Bharat's IndicWhisper; 18.3 to 59.9 for Google; 15.2 to 42.3 for Azure, moving from studio-read to Gramvaani speech	Channel and dialect together cost a lot of accuracy

Three cautions apply. The Hindi-English corpora are prompted sentences, and a caller describing a grievance speaks differently. The systems date from 2018 to 2023. And the Vistaar jump is not code-switching: the paper describes Gramvaani as telephone-quality speech focused on regional and dialectal variation of Hindi, so channel and dialect are tangled together.

What survives is narrower and still useful. A department that tests only clean, monolingual Hindi has not tested its helpline.

Four Places Code-Switched Calls Break

1. The audio channel

The standard telephone codecs PCMU and PCMA, defined in ITU-T G.711, carry audio at 8,000 samples a second, one channel, eight bits per sample, according to the IETF's RFC 3551. A signal sampled at 8,000 Hz holds nothing above 4,000 Hz. That is a property of the sampling, not of any vendor.

Some speech models are trained on 16 kHz audio. The Vistaar team resampled all audio to 16 kHz for its IndicWhisper models. An 8 kHz call fed into such a model is stretched to the expected rate, which adds samples and does not restore the missing frequencies. The Bhashini documentation says the sampling rate and format a client declares must match the audio actually recorded, so declare 8 kHz honestly. Our integration guide covers the call sequence.

2. The output script

A model that writes both Hindi and English must choose labels for both. In the HingCoS experiments the unified character set had 95 targets: 26 English characters, 68 Hindi characters and a word separator. The authors list "more than double expansion in the target set" and "requirement of more data for reliable modelling" among the challenges, and say little code-switching data exists.

For a helpline the visible effect is the transcript. "Application" may arrive in Roman letters, in Devanagari, or as a mixture within one complaint. The portal's search, duplicate check and officer view all handle those differently. Decide the department's rule first: which script is stored, and whether English terms are kept as spoken.

3. How errors are counted

Word error rate treats any spelling difference as a mistake. AI4Bharat's 2026 paper on orthographically-informed evaluation (Bhogale et al., arXiv 2603.00941) notes that many code-mixed words show five or more spellings. Scoring that accepts valid spellings lowered measured error rates by an average of 6.3 points across languages. For Hindi the average difference between standard and spelling-aware scoring was 5.8 points, which the authors attribute to the large number of loaned words. These numbers come from the IndicVoices benchmark and are not telephone-specific, but the lesson transfers.

Without an agreed spelling rule, a bake-off rewards the engine whose output happens to match the transcriber's habits. A correct reading of "complaint" in the wrong script scores as an error, and a wrong word in the right script may not.

4. Numbers and names

A grievance call is mostly names, scheme titles, amounts, dates and reference numbers. AI4Bharat's IndicContextEval benchmark measures how well speech models handle named entities, and tests entity lists supplied in both English and native script. The lesson is to give entities their own score. The orthography paper also treats numbers, dates and currencies as places where several written forms are valid. If a caller says a nine-digit complaint number half in Hindi and half in English, the engine has to carry the digits across the switch. Score that case on its own.

A Test Deck for Code-Switched Calls

The unit of testing is a recording set from the department's own line, held out from any tuning. Build it in five groups so a weak spot is not averaged away.

Group	What the caller does	What to score
A. Plain Hindi (or the local language)	Describes a complaint with no English	Word and character error rate; the baseline
B. English nouns inside Hindi	Names a document, scheme or department in English mid-sentence	Capture of the English term in the agreed script
C. Numbers across the switch	Gives a complaint or application number in mixed Hindi and English digits	Exact-match rate of the full number; read-back success
D. English with local words	Speaks mostly English with Hindi phrases	Same engine setting as the caller's menu choice; handoff rate
E. Line conditions	Calls from a mobile in a market, a vehicle or a weak-signal village	Group A to D scores, repeated; report separately

Agree three rules in writing before running anything. First, the spelling policy: which scripts and spellings count as correct for English terms. Second, the entity list: scheme names, department names and document types the portal must capture. Third, the pass mark, set from the baseline measured on the department's current line. No honest target exists before that baseline does.

Then read failures against likely causes:

Symptom	Likely cause	First thing to try
English words garbled, Hindi fine	The configured language model is weak on English insertions	Test another engine for that language
Correct meaning, high error rate	Spelling variance counted as error	Re-score with the agreed spelling rule
Same engine passes in the office, fails on mobiles	Channel mismatch or wrong declared sampling rate	Check the declared rate and the codec path
Complaint numbers wrong, text fine	Digits across the language switch	Add number read-back before filing
Scheme names inconsistent in the portal	No script policy	Fix the stored-form rule

What Changes in the Helpline's Workflow

Today the caller presses a number for a language, an agent who may or may not speak it listens, and the agent types the complaint into the portal. Code-switching is handled by the agent's ear.

With a tested voice layer, the caller's menu choice sets the engine's language. The engine transcribes, the system reads the complaint number back before filing, and a failed read-back moves the call to a human in the same language where possible. The portal stays the record, and named officers keep every decision. A weekly report shows, per test group, the share of calls handed to a person, so a weak group is visible instead of buried in an average.

Roadmap

Weeks 1 to 2. Pick one district and no more than three languages. Use real calls with the proper authority, or record test calls on the same carrier paths. Check retention rules first; our guide to grievance data retention sets out what may be kept. The DPDP Act's consent and penalty sections commence on 13 May 2027, and the Act is only partly in force today.
Weeks 3 to 4. Transcribe, write the spelling policy and entity list, and run each candidate engine on groups A to E. Our engine comparison guide covers choosing per language.
Weeks 5 to 6. Set the pass mark from the baseline, select an engine per language, and decide which groups need number read-back or human handoff.
After go-live. Re-run the deck whenever an engine version, a telecom provider or a district changes.

Risks and Mitigation

Testing on a studio microphone. The result describes the demo, not the line. Record on the real call path.
One average for all groups. A passing overall score can hide group C, the one that writes the wrong number into a grievance.
Borrowing a research score as a promise. A 2018 corpus result is not a service level. Do not copy it into a tender.
No spelling rule. The vendor's script choice wins by default. Write the rule first.
Silent model updates. Service IDs point to named models, so an update can change behaviour with no code change.

Key Takeaways

Callers mix languages mid-sentence and the line delivers 8 kHz audio, so a clean monolingual demo does not test a helpline.
The Hindi-English telephone corpora we found are recorded at 8 kHz, but they are read sentences scored on older systems, and none ranks current engines.
Four failure points deserve their own tests: the channel, the output script, the way errors are counted, and numbers and names.
Build a five-group test deck from the department's own recordings and set the pass mark from the current line's baseline.
Read back every complaint number before filing it.

Conclusion

The honest summary of the evidence is that nobody can tell a helpline cell in advance how an engine will handle its callers' mixed Hindi and English on a telephone line. What the cell can do is record its own calls, agree how errors will be counted, test the cases that hurt, and launch each language only after it passes.

Government leaders exploring AI-powered citizen engagement can begin with a focused pilot in one department or constituency to validate impact before scaling statewide. Aisewak helps public institutions deploy multilingual Voice AI solutions designed specifically for Indian governance. The multilingual voice agent page describes the language coverage and acceptance testing approach.

FAQ

What is code-switching in a helpline call?

It is the use of two languages within one sentence or conversation. On an Indian helpline this is usually Hindi with English nouns, such as a document, scheme or department name, or a regional language with English terms.

Do speech engines handle Hinglish well?

We found no published study that scores a current commercial or Bhashini engine on spontaneous Hindi-English telephone speech, so the department should test vendor claims on its own recordings.

Why does 8 kHz audio matter?

Standard telephone codecs sample at 8,000 Hz, which cannot carry content above 4,000 Hz. Some speech models, including the IndicWhisper models in the Vistaar benchmark, take 16 kHz input, so call audio is stretched to fit and the engine must be tested on it.

How much worse is telephone speech than studio speech?

In the Vistaar benchmark, one Hindi model's error rose from 7.6 to 26.8, and Google's from 18.3 to 59.9. That comparison mixes telephone quality with dialect variation, so read it as a warning and not as a precise penalty.

Why can a correct transcript score as an error?

Word error rate counts spelling differences as mistakes. Code-mixed words often have several accepted spellings, so a department should agree a spelling policy before comparing engines.

Should the transcript be in Devanagari or Roman script?

The department decides, based on how its portal searches and displays text. Choose one stored form for English terms and test engines against it.

How should complaint numbers be handled?

Read the number back to the caller before filing, and score number capture as its own test group. A transcript can read well and still carry a wrong digit.

Can this be tested before any contract is signed?

Yes. A sample of recorded calls from one district, run through each candidate engine, is enough to compare them. Check retention and consent rules for the recordings first.

Schema Markup Suggestions

Article: headline, description, datePublished and dateModified of 2026-10-16, publisher AiSewak.
FAQPage: the eight Q&A pairs above.
HowTo (optional): the five-group test deck and the four-step roadmap.

Suggested External References

Sreeram, Dhawan and Sinha, Hindi-English Code-Switching Speech Corpus — arXiv 1810.00662, 2018
Dhawan et al., Investigating Target Set Reduction for End-to-End Speech Recognition of Hindi-English Code-Switching Data — arXiv 1907.08293, 2019
Diwan et al., MUCS 2021: Multilingual and Code-Switching ASR Challenges for Low Resource Indian Languages — arXiv 2104.00235
Bhogale et al., Vistaar: Diverse Benchmarks and Training Sets for Indian Language ASR — arXiv 2305.15386, 2023
Bhogale et al., Towards Orthographically-Informed Evaluation of Speech Recognition Systems for Indian Languages — arXiv 2603.00941, March 2026
Joshi et al., IndicContextEval — arXiv 2606.19157, Interspeech 2026
IETF, RFC 3551, RTP Profile for Audio and Video Conferences with Minimal Control — datatracker.ietf.org
Bhashini APIs documentation, Digital India Bhashini Division — dibd-bhashini.gitbook.io
Digital Personal Data Protection Act, 2023 — commencement of consent and penalty provisions on 13 May 2027

Callers to Indian helplines mix Hindi and English mid-sentence, over 8 kHz phone audio. We found no published score for current engines on that. A guide to the four things that break and the five-group test a state helpline cell should run first.

LinkedIn Executive Summary

A state helpline cell that wants callers to speak naturally has a problem the demo never shows: real callers mix Hindi and English inside one sentence, and the audio arrives at 8 kHz.

The research that exists is useful and limited. Hindi-English telephone corpora were recorded at 8 kHz, but they are read sentences scored on older systems. AI4Bharat's Vistaar benchmark shows Hindi error rising sharply on telephone-quality speech, though that test mixes channel and dialect. A 2026 AI4Bharat paper shows that code-mixed words often have five or more valid spellings, so a raw error rate can mislead.

The practical answer is a test deck built from the department's own recorded calls: plain speech, English nouns inside Hindi, numbers across the switch, English-dominant speech and difficult line conditions. Agree the spelling policy and entity list first, set the pass mark from the current line, and read back every complaint number before filing it.

AI Search Optimization Summary

Entities: Bhashini, Digital India Bhashini Division, MeitY, AI4Bharat, IIT Guwahati, HingCoS corpus, MUCS 2021, Vistaar, IndicVoices, IndicContextEval, Gramvaani, ITU-T G.711, RFC 3551, DPDP Act, Aisewak

Topics: Hindi-English code-switching speech recognition, Hinglish ASR on telephone audio, 8 kHz narrowband speech recognition, orthographic variation in word error rate, helpline speech recognition test design, complaint number capture

Semantic keywords: code-mixed speech recognition India, Hinglish telephone corpus, 8 kHz G.711 PCMU ASR, loaned words spelling variation, OIWER, entity capture across language switch, helpline acceptance testing, script policy for English terms in Devanagari transcripts