Skip to content

Systems and records

Business call code switching dataset: scoping English-led calls

By SourceX Editorial · Reviewed by Noah Loul ·

10 sections

Short answer

Scope English-led business calls by measured language share, not by the phone system's language flag. Label in separate layers on the audio timeline: speaker turns, language segments with BCP 47 tags, typed switch events, repair turns, transfers and verified outcomes from systems of record. Compute share by time and by tokens, per call and per speaker, and split calls at transfers before deciding what enters the dataset.

Call to transcript

Business call code switching

This page covers business call code switching dataset: scoping English-led calls.Illustrative

Key takeaways

  • A routing or CRM language field records a choice made before the call, not the language mix of the conversation.
  • Measure English share by speech time and by tokens, per call and per speaker, excluding names, numbers and codes.
  • Type every switch, because within-sentence switches, single inserted terms and discourse markers fail in different ways.
  • Split a call at a transfer to a bilingual agent and compute share for each span.
  • Attach outcomes from systems of record after a maturity window, never from what the caller said.

How should English-led business calls with code-switching be scoped and labeled?

Scope them by measured language share, not by the language field in the phone system, and label them in layers that never overwrite each other: who spoke when, which language each stretch of speech is in, where and why the speaker switched, and what verifiably happened after the call. Each layer carries timestamps on the original audio so any label can be traced back to the seconds that justify it.

English-led means most of the talk is in English, with the caller or the agent moving into another language for a name, an address, a product term, a clarification or a whole stretch of explanation. These calls are common in U.S. customer service, field service dispatch, insurance claims intake and healthcare front desks, and they are exactly where speech systems tuned on monolingual audio lose words.

The asset on this page is a language share and switching annotation plan: seven layers with their fields and checks, a way to compute share per call and per speaker, a switch taxonomy, and rules for attaching outcomes without letting the conversation itself decide them.

Why does a routing language flag tell you so little about the call?

The language recorded by an IVR menu or a CRM contact field records a choice made before anyone spoke. A caller who pressed 1 for English may describe a symptom in Spanish; an agent on an English queue may switch to Tagalog for a family member who joins the line. Selecting calls by that flag yields a set whose real language mix nobody has measured.

Automated tools do not close the gap on their own. The Whisper README explains that transcription runs over a sliding 30-second window, and its language detection example pads or trims audio to 30 seconds and returns one probability per language for that clip. That is one language guess per window, which is coarse for a call where a switch can last two words. The Whisper model card also notes uneven accuracy across languages, accents and dialects, and that outputs may include text that was not actually spoken.

So automatic language identification is a useful first pass for routing audio to annotators, but the language share that defines the dataset should come from human-checked segments.

What should the language share and switching annotation plan include?

Seven layers, each with its own owner and check. The table lists the fields that make each layer usable as delivered; teams can add fields, but removing one usually breaks a join or an audit.

Run annotation in two passes. The first pass fixes speaker turns and language segments; the second, done by a different annotator fluent in both languages, types the switches, marks repairs and reviews any segment where automatic and human language tags disagreed. Splitting the work this way keeps the switch labels from inheriting a mistake in the turn boundaries.

Measure agreement on a shared sample before scaling. Two annotators should label the same calls independently, and the plan should state how close their segment boundaries must be, in milliseconds, and how often their switch types must match. Disagreements usually trace back to glossary gaps, such as whether a brand name or a common loanword counts as a switch, and fixing the glossary early is cheaper than relabeling later.

What should the language share and switching annotation plan include?
LayerUnitRequired fieldsAcceptance check
Call headerOne callcall_id, recording rights basis, channel layout (mono or dual), duration, queue, transfer flagRights basis recorded before any audio leaves the source system
Speaker turnsTurnspeaker_id, role (agent, customer, interpreter, other), start and end timeDiarization reviewed by a person on every call with a switch
Language segmentsStretch of speech in one languageBCP 47 tag, start and end time, method (human or automatic), annotator idSegments tile each turn with no gaps or overlaps
Switch eventsPoint where language changesposition, direction, type, trigger, speaker_idEvery boundary between segments of one speaker has an event
Repair turnsTurnrepair type (repeat, rephrase, spell, translate), what it repairedLinked to the turn it repairs
Transfer contextHandofftarget (bilingual agent, interpreter line, queue), time, reason if statedCall split into pre- and post-transfer spans
Verified outcomeOne call or caseoutcome label, source system, record id, date observed, maturity dateOutcome comes from a system of record, not from the transcript

How do you measure language share per call?

Compute share two ways and keep both: by speech time, from the language segment timestamps, and by tokens, from the transcript. Time-based share reflects what an audio model hears; token-based share reflects what a text model reads. They diverge when one language is spoken faster or when long pauses fall inside one language's segments.

At a glance

How do you measure language share per call

Exclude language-independent tokens from the token count: personal and company names, numbers, product codes, email addresses and acronyms that read the same in either language. Without that rule, a customer spelling a policy number inflates whichever language the annotator chose for digits.

Report share per speaker as well as per call. A call where the agent speaks only English and the customer speaks mostly Spanish can look English-led overall because agents talk more, while the customer side, which is often what the buyer wants to model, is not English-led at all.

  • english_share_time: English speech seconds divided by all labeled speech seconds, per call and per speaker.
  • english_share_tokens: English tokens divided by all language-tagged tokens after exclusions.
  • switches_per_minute: count of switch events divided by call minutes, per speaker.
  • longest_non_english_span: the duration of the longest continuous non-English stretch, which flags calls that are English-led overall but contain long monolingual passages.
  • share_band: the buyer's own bands, for example mostly English, mixed and mostly other language, with cut-offs written in the specification before annotation starts.

Which switches should annotators mark, and how?

Mark every change of language within a speaker's speech, and type it, because the types fail differently in recognition and comprehension. Tag languages with BCP 47 tags, the IETF convention defined by RFC 5646, in which a tag is built from subtags such as a language and an optional region, so es-MX and es-US can be told apart when the buyer cares.

A short taxonomy keeps annotators consistent:

  • Between sentences: the speaker finishes a sentence in one language and starts the next in another.
  • Within a sentence: the language changes mid-clause, the hardest case for recognition and the one most worth sampling.
  • Single inserted word or phrase: one term, often a product name, a dish, a place or a kinship word, dropped into the other language.
  • Tag or discourse marker: short particles such as okay, pues or sige that carry little content but signal a switch.
  • Established loanword: a word fully adopted into English; label it English and log it in a project glossary so the decision is applied the same way every time.
  • Ambiguous token: a word that is valid in both languages in context; give it a project-defined ambiguous label rather than forcing a choice.

How do you check switched segments against the audio and outcomes against systems of record?

Check the audio before trusting either. A transcript that reads cleanly can still be misaligned by seconds, can attribute the customer's words to the agent, or can have smoothed a switched phrase into the dominant language. Sample switched segments, play the audio, and confirm the words, the speaker and the language tag together.

Dual-channel recordings, with agent and customer on separate channels, make speaker attribution far more reliable than mono. For mono audio, a diarization toolkit such as the open-source pyannote.audio gives a starting segmentation, but a person should review turns around every switch, because a wrong speaker label silently moves a switch from customer to agent.

Outcomes need the same discipline. Whether a claim was approved, a repair was completed on the first visit or an order was kept comes from the system that records it, matched by record id, after a maturity window the buyer sets. A customer saying thanks, that fixes it, is a conversational signal; the closed ticket with no repeat call within the window is the outcome. Store the two in different layers so a model is never trained to predict a label that was inferred from its own input.

How is a warranty call with two switches and a transfer labeled? (illustrative)

This example is illustrative and not drawn from any real call. A customer calls an appliance warranty line on the English queue. The call lasts eleven minutes. For the first four minutes both speakers use English; the customer then describes a burning smell in Spanish for about forty seconds, the agent asks in English for the model number, and the customer reads it out with digits in English.

Annotation produces two switch events for the customer (English to Spanish within a sentence, Spanish back to English between sentences), a repair turn where the agent rephrases a question, and a language-independent span for the model number that is excluded from token share. At minute six the agent transfers the customer to a bilingual specialist, who continues mostly in Spanish.

The transfer is the exception that changes scope. Measured over the whole call, English share is near the middle band; measured before the transfer, the call is clearly English-led, and after it, mostly Spanish. The plan handles this by splitting the call at the transfer into two spans with their own shares, so the English-led span can enter the dataset and the other span can be routed to a different subset or left out. The verified outcome, a repair order closed eight days later with no repeat call in the window, attaches to the call_id, not to either span's text. Recording both spans, even when only one is delivered, lets a later reviewer see why the call was cut where it was.

What can public code-switching corpora teach, and what can't they?

Public corpora are useful for annotation conventions and for benchmarking language identification, but none of them is a set of business calls with verified outcomes. Use them to calibrate annotators and tools, then specify your own data for the job.

The Bangor Miami corpus on TalkBank holds 56 recordings and transcripts of informal conversation among 84 speakers living in Miami. The SEAME corpus distributed by the Linguistic Data Consortium contains spontaneous Mandarin-English speech recorded in Singapore and Malaysia in conversational and interview styles, and the catalog lists language boundary detection and language identification among its intended uses. LinCE gathers ten text corpora across four language pairs into one benchmark with tasks that include language identification, named entity recognition, part-of-speech tagging and sentiment analysis.

What these sets lack is the business context: queues, transfers, agents following scripts, domain vocabulary and a record of what happened next. They also differ in license terms, so check each corpus's terms before using it even for calibration.

When should a code-switching scope narrow or go to specialists?

Narrow the scope, or bring in a specialist, when any of the conditions below holds. Each is a point where more volume makes the dataset worse rather than better.

  • The recording consent covered quality assurance but not use for training AI; the rights basis for training is a separate question, and counsel reviews it on the specific facts.
  • Annotators are not fluent in the other language or its regional variety; switches will be mislabeled as noise or as the dominant language.
  • Calls involve minors, health details, payment card numbers read aloud or other sensitive content that de-identification must remove from audio as well as transcript.
  • The other language appears mostly through an interpreter line; interpreted speech is a different phenomenon from a speaker's own switching and needs its own subset.
  • Diarization error is high on mono recordings, so per-speaker share cannot be computed reliably without manual turn review on every call.
  • The outcome window has not closed; attach outcomes only after the maturity date in the specification.

How does the plan become a data request?

A request built on this plan names the language pairs and regional varieties, the share bands and how they are computed, the switch taxonomy, the channel layout, the outcome types with their systems of record and maturity windows, and how transfers are split.

Call datasets like this are not held in stock; SourceX sources them from U.S. businesses on request and carries the selected recordings through rights review, contracting and preparation. Share bands and a switch taxonomy belong in the request-data form next to scope, volume, language, format, timing and licensing needs.

Frequently asked questions

What is code-switching in a business call?

Code-switching is a speaker moving between two languages within one conversation, sometimes within one sentence. In business calls it often happens for names, addresses, product terms or to explain something more precisely. It differs from a call handled through an interpreter, where a third person translates.

Can automatic language identification label switches on its own?

It is a good first pass but not a reliable final label. Tools that assign one language per window of audio miss short switches, and recognition accuracy varies across languages and accents. Human review of switched segments should set the dataset's language share.

Should language share be measured by time or by words?

Measure both and keep both. Time-based share matches what an audio model hears, while token-based share matches what a text model reads. Exclude names, numbers and codes from token counts so spelled-out identifiers do not skew the result.

How should a call with a transfer to a bilingual agent be handled?

Split it at the transfer and compute share for each span. The span before the transfer may be English-led while the span after is mostly the other language. Attach the verified outcome to the call id rather than to either span.

Why keep outcomes separate from what was said on the call?

A customer sounding satisfied is not the same as a resolved case. Outcomes should come from the system that records them, after a set maturity window. Keeping them in their own layer prevents training a model to predict a label inferred from its own input.

Sources

  • The Whisper README says transcription processes audio with a sliding 30-second window, and its language detection example pads or trims audio to 30 seconds and returns probabilities per language. Source
  • The Whisper model card reports uneven performance across languages and across accents and dialects, and says predictions may include text that was not actually spoken. Source
  • RFC 5646 (BCP 47) defines language tags composed of one or more subtags, each refining the language identified, such as language and region subtags. Source
  • pyannote.audio is an open-source Python toolkit for speaker diarization with pretrained pipelines. Source
  • The Bangor Miami corpus contains 56 recordings and transcripts of informal conversation involving 84 speakers living in Miami, Florida. Source
  • The LDC catalog describes SEAME as spontaneous Mandarin-English speech recorded in Singapore and Malaysia in conversational and interview styles, intended for language boundary detection, language identification and code-switched speech recognition. Source
  • LinCE combines ten code-switching corpora across four language pairs, with tasks including language identification, named entity recognition, part-of-speech tagging and sentiment analysis. Source

Tell us what data you need

Share scope, volume, format, timing and licensing needs. Datasets are sourced on request.

See if your company qualifies