Can AI Transcription Reliably Handle Multilingual Interviews?

Can AI Accurately Transcribe Interviews That Use More Than One Language?

You interview Filipino teachers about their experiences using artificial intelligence in the classroom. Although the interview questions are written in English, participants frequently answer in Filipino, switch between Filipino and English, and occasionally use expressions from their regional languages.

You upload the recordings to an AI transcription tool that advertises multilingual support. The resulting transcript looks readable, but some Filipino expressions have disappeared, certain English technical terms are incorrectly recognized, and a few statements appear to have been translated rather than transcribed.

Can AI reliably process this kind of interview? More importantly, how can researchers establish whether the resulting transcript faithfully represents what participants said across languages?

AI Can Transcribe Multiple Languages, but Code-Switching Requires Particular Care

How Multilingual AI Transcription Works and Where It Can Fail

What Counts as a Multilingual Research Interview?

A multilingual interview involves the use of more than one language during the research interaction. This may occur because the interviewer and participant use different languages, because participants switch languages during their responses, or because the research setting involves several linguistic communities.

Not every multilingual interview has the same structure. One participant might complete an entire interview in Filipino, while another alternates between Filipino and English within individual sentences.

These differences matter because recognizing several languages separately is not necessarily equivalent to recognizing spontaneous language switching.

What Is the Difference Between Multilingual Speech Recognition and Code-Switching?

Multilingual speech recognition refers to the ability of a system to recognize spoken content in more than one language. Code-switching refers to the alternation of languages within a conversation or utterance.

Consider a participant who says, "Mahirap gamitin noong una, but eventually I became comfortable with it."

A multilingual system might recognize Filipino and English successfully when they occur in separate recordings but struggle when both languages appear in one sentence.

Code-switching therefore introduces an additional recognition challenge rather than merely increasing the number of supported languages.

What Does Published Research Show About Code-Switching?

Research has repeatedly identified code-switching as a challenging condition for automatic speech recognition.

Sitaram and colleagues (2019), in A Survey of Code-switched Speech and Language Processing, reviewed computational approaches to mixed-language speech and text. They highlighted the importance of code-switching for multilingual communities and the limitations of available linguistic resources for many language combinations.

Mustafa and colleagues (2022) conducted a review of code-switching in automatic speech recognition. Their review examined 42 experimental papers and found substantial variation in performance across systems and language combinations. The authors identified challenges involving pronunciation differences, acoustic modeling, language modeling, and the availability of suitable datasets.

More recently, Agro and colleagues (2026) published a systematic literature review of code-switching in end-to-end automatic speech recognition. Their work examined published research on language combinations, datasets, evaluation metrics, and model approaches, highlighting continuing methodological and resource gaps.

These studies support the conclusion that multilingual recognition and code-switching performance cannot be assumed from general speech recognition capabilities. They do not establish a universal accuracy percentage for multilingual research interviews or for any particular language pair.

Why Does AI Sometimes Recognize One Language Better Than Another?

Multilingual speech recognition performance depends partly on the languages represented in the model's training data and the diversity of speech patterns included in those recordings.

Languages with substantial available training resources may receive stronger recognition support than languages with relatively limited speech datasets. However, dataset size alone does not determine performance. Recording quality, linguistic characteristics, model architecture, and evaluation conditions also matter.

Radford and colleagues (2023), in their work introducing Whisper, demonstrated multilingual speech recognition capabilities using large-scale weak supervision. Their evaluation also showed considerable variation in performance across languages and datasets.

Consequently, a tool's claim to support numerous languages should be interpreted as a statement of capability, not evidence that all supported languages are transcribed equally accurately.

The researcher must establish whether the selected system performs adequately for the languages and speaking conditions represented in the study.

What Happens When AI Detects the Wrong Language?

Some transcription systems automatically identify the language spoken in a recording. Others require users to specify the language before transcription begins.

Automatic language identification may be useful when the recording contains one dominant language. However, difficulties can arise when participants switch languages frequently or use short expressions from another language.

For example, a system may identify an interview as English because most questions are asked in English. It may then process Filipino responses less accurately, depending on how the system handles mixed-language input.

Alternatively, a system might misinterpret an unfamiliar phrase as acoustically similar words from the dominant language.

Researchers should test whether automatic language detection, manual language selection, or another supported configuration produces better results for their recordings. No single setting is necessarily optimal across all multilingual systems.

Can AI Preserve Filipino-English Code-Switching?

Potentially, yes. Some multilingual speech recognition systems can represent Filipino and English words within the same transcript. However, their actual performance on spontaneous Filipino-English code-switching must be established through testing.

Consider the following hypothetical statement:

Participant: "Ginagamit ko naman ang AI sa lesson planning, pero I still check the accuracy of the content."

A faithful original-language transcript should preserve the Filipino and English expressions that were actually spoken.

If the system instead produces an entirely English sentence, the researcher must determine whether the tool performed translation, normalization, or incorrect recognition.

These outcomes have different methodological implications. A translated sentence may preserve much of the intended meaning while no longer representing the participant's original wording. An incorrect recognition may introduce information that the participant never expressed.

Researchers should avoid assuming that a fluent English output is necessarily an accurate transcription of a Taglish interview.

Is Transcribing a Multilingual Interview the Same as Translating It?

No. Transcription and translation are distinct processes, even when a software platform offers both.

The translated version may be useful for analysis by researchers who do not understand the original language. However, it introduces interpretive decisions that should not be confused with word recognition.

Researchers considering simultaneous AI transcription and translation should evaluate both processes separately.

Why Can Language Switching Affect Participant Meaning?

Participants may switch languages for reasons that extend beyond convenience. A language change may accompany emphasis, a culturally specific expression, a change in conversational stance, or a distinction that is easier to communicate in one language.

However, researchers should not automatically assign a particular psychological or social meaning to every language switch. The significance of code-switching depends on the context, participant, and research question.

If an AI transcription system removes the original-language expression or replaces it with a standardized translation, information relevant to interpretation may disappear.

For example, a Filipino participant might describe a teaching experience using the word "nakakahiya." Depending on context, the expression may involve embarrassment, shame, or concern about social evaluation. Automatically replacing it with one English word may obscure nuances that require interpretation.

This is why transcription errors affecting participant meaning deserve particular attention in multilingual qualitative research.

How Should Researchers Evaluate Multilingual Transcription Accuracy?

A meaningful evaluation should use recordings that reflect the language combinations, speaking styles, and recording conditions present in the study.

Researchers can prepare reference transcripts verified by appropriately qualified speakers and compare them with AI-generated transcripts.

Word error rate may be useful, but multilingual evaluation requires consistent decisions about word segmentation, spelling conventions, and the treatment of code-switched expressions.

For languages with different writing systems or segmentation conventions, character error rate or other complementary measures may be appropriate.

Researchers should also examine errors qualitatively. A single incorrectly recognized culturally specific expression may matter more to interpretation than several minor spelling differences.

When feasible, accuracy should be examined separately for each language and for code-switched passages rather than relying only on an overall score.

Who Should Verify Multilingual Research Transcripts?

Verification requires more than familiarity with the topic. Reviewers should have sufficient linguistic competence to understand the languages and relevant varieties used in the recordings.

For Filipino-English interviews, this may involve researchers or assistants who can understand both languages and recognize common patterns of code-switching.

Where regional languages are involved, additional linguistic expertise may be necessary. A reviewer who understands Filipino and English should not automatically be assumed capable of verifying every Philippine language.

Researchers should also consider cultural and contextual knowledge when interpreting ambiguous expressions, while ensuring that corrections remain grounded in the recording.

The broader requirement for manual verification of AI-generated transcripts remains relevant, but multilingual interviews may require additional linguistic expertise.

Should the Original-Language Transcript Be Preserved?

Preserving a verified original-language transcript is generally advisable when multilingual speech contributes to the research evidence.

The original-language version provides a reference for checking translations, reviewing interpretations, and examining expressions that may not have direct equivalents in another language.

Researchers may create a separate translated transcript for analysis or publication when appropriate. The relationship between the original and translated versions should remain traceable through consistent identifiers or timestamps.

All versions should be managed according to the study's confidentiality, consent, and data protection requirements.

The goal is to make multilingual evidence accessible for analysis without silently replacing the participant's original language with the researcher's preferred representation.

Transcribing Filipino-English Interviews About AI in Education

The AI-generated English output removes the participant's qualification and changes the relationship between usefulness and trust. The resulting statement suggests confidence in AI-generated material, whereas the original expresses caution.

This example illustrates why a multilingual transcript should not be evaluated solely by whether the final English text sounds natural. Fluency can conceal both recognition and translation errors.

Common Misunderstandings About Multilingual AI Transcription

How Should You Plan AI Transcription for Multilingual Research?

Multilingual transcription should be planned around the languages participants actually use, not merely the language in which the interview protocol was written.

If participants may respond in multiple languages, establish whether the transcription system can preserve those languages and whether the research team has the expertise to verify the resulting records.

For multilingual research, a defensible transcript verification workflow should explicitly address language switching, translation boundaries, and the qualifications of reviewers.

The objective is not to force multilingual speech into a single standardized language. It is to preserve the evidence participants provided while making that evidence sufficiently accurate and accessible for analysis.

Before Using AI Transcription for Multilingual Interviews

Frequently Asked Questions About Multilingual AI Transcription

Multilingual Transcription Must Preserve the Languages Participants Actually Use

How to Cite This Guide

This guide is intended to be read, shared, and used in research, teaching, and academic work. If you draw on its ideas, explanations, or other content, please acknowledge the source by citing the guide. Doing so gives appropriate credit and helps your readers locate the original resource.