Can a computer discover connections between languages simply by listening to people speak? A fascinating research study from IIT Kharagpur explored this question, revealing interesting similarities between Indian languages and even some unexpected connections.
India is home to an extraordinary variety of languages. From Kannada and Malayalam in the south to Hindi, Gujarati, Bengali and Kashmiri in other regions, each language carries its own history, identity and cultural traditions.
But have you ever wondered how similar two languages might sound to a computer, even when they are written differently or spoken thousands of kilometres apart?
Researchers Debapriya Sengupta and Goutam Saha from the Indian Institute of Technology (IIT) Kharagpur investigated this question using speech-processing technology. Their research explored how computers could identify similarities between languages by analysing spoken audio rather than relying only on written words.
The study, published in 2015, offers an interesting example of how artificial intelligence and speech technology can help researchers investigate relationships between languages.
1. Why Are Some Indian Languages Similar to Each Other?
Languages do not develop in isolation. They evolve over centuries through migration, trade, cultural exchange, geographical contact and historical events.
One important concept in linguistics is the language family. A language family consists of languages that developed from a common ancestral language.
For example, Kannada, Tamil, Telugu and Malayalam belong to the Dravidian language family. Hindi, Bengali, Gujarati and Marathi belong to the Indo-Aryan branch of the Indo-European language family.
Languages from the same family may preserve similarities in vocabulary, grammar, pronunciation and other linguistic features. Neighbouring languages can also influence each other through everyday communication and cultural contact.
However, the relationship between languages is not always straightforward. Two languages can share some characteristics without being closely related, and closely related languages can develop very different pronunciations over time.
This creates an important question for researchers: Can a computer identify similarities between languages by examining how they sound?
The IIT Kharagpur researchers designed an experiment to investigate precisely this question.
2. The Research Paper Behind the Discovery
The study was titled “Study on Similarity among Indian Languages Using Language Verification Framework.”
It was authored by Debapriya Sengupta and Goutam Saha and published in the journal Advances in Artificial Intelligence in 2015.
Instead of comparing only written text, the researchers developed computer systems that analysed spoken language recordings.
Their database contained recordings from 16 languages:
Assamese
Bengali
English
Gujarati
Hindi
Kannada
Kashmiri
Konkani
Malayalam
Marathi
Odia
Punjabi
Sanskrit
Tamil
Telugu
Urdu
The recordings came from All India Radio news bulletins. The researchers selected this source because it provided relatively clear speech, multiple speakers and a variety of spoken content.
They prepared approximately four hours of clean training speech for each language. The testing data included recordings lasting 3 seconds, 10 seconds and 30 seconds.
This allowed them to investigate whether a computer could identify a language from a short audio sample and whether longer recordings improved its ability to distinguish between languages.
3. How Did the Computer Identify Similarities Between Languages?
The researchers used a technique known as a language verification framework.
Imagine training a computer to recognise Kannada. You provide it with many examples of Kannada speech so that it learns characteristics associated with that language.
Next, you play it a recording of another language, such as Malayalam.
The computer attempts to determine whether the new recording resembles the Kannada speech it learned during training.
If the computer mistakenly accepts the Malayalam recording as Kannada, that mistake becomes a useful measurement for the researchers.
Why? Because repeated confusion between two languages may indicate that they share certain acoustic characteristics.
The researchers used this idea to investigate similarities among all 16 languages.
Understanding false acceptance
In language verification, a false acceptance occurs when a system accepts a recording from a different language as belonging to its target language.
For example:
The computer is trained to recognise Language A.
A speaker provides a recording in Language B.
The computer incorrectly accepts Language B as Language A.
Researchers record this error and examine whether it happens repeatedly.
One mistaken classification does not prove that two languages are similar. The researchers therefore tested multiple recordings and used four different verification systems to reduce the influence of unusual results from any one system.
Their approach transformed a computer's mistakes into a way of investigating potential linguistic relationships.
4. What Technology Did the Researchers Use?
The study combined two types of speech-feature extraction techniques with two types of statistical machine-learning models.
Although the technical names may sound complicated, their basic functions are understandable.
MFCC: Examining the sound of speech
MFCC stands for Mel Frequency Cepstral Coefficients.
When people speak, their voices produce sound waves with different frequencies and patterns. MFCC converts parts of those sound patterns into numerical features that a computer can analyse.
Think of it as converting a voice recording into a mathematical description of its sound characteristics.
SFCC: Another way to describe speech sounds
SFCC stands for Speech Signal Based Frequency Cepstral Coefficients.
This method also extracts numerical features from speech, but it uses a filter bank derived from characteristics of the speech signal.
The researchers tested SFCC alongside MFCC to investigate whether different methods of representing speech would produce useful results.
SDC: Capturing changes over time
SDC stands for Shifted Delta Coefficients.
Speech is not simply a collection of individual sounds. The way sounds change and follow one another also carries information.
SDC captures patterns of change across time, providing additional information about the structure of spoken language.
GMM and SVM: Learning to distinguish languages
The researchers used two statistical modelling approaches:
Gaussian Mixture Model (GMM): Represents the distribution of speech features using a combination of statistical distributions.
Support Vector Machine (SVM): Learns a decision boundary that helps separate one category of data from another.
By combining the two feature-extraction methods with the two modelling methods, the researchers created four verification systems.
They then compared their results rather than relying on a single system.
It is important to note that this was a speech-processing and machine-learning study conducted in 2015, not a modern large language model comparing the meanings of sentences. Its primary evidence came from acoustic patterns in spoken recordings.
5. Why Did the Researchers Test 3-Second, 10-Second and 30-Second Recordings?
One of the interesting aspects of the experiment was its use of different recording lengths.
Imagine hearing someone speak for only three seconds. You might recognise their language, but the short recording may not contain enough information to make a reliable distinction.
A 30-second recording provides a much larger sample of pronunciation, sound patterns and speech variation.
The researchers observed that the average Equal Error Rate (EER) generally decreased as the test recording length increased from 3 seconds to 10 seconds and then to 30 seconds.
EER measures the point at which the false-acceptance rate and false-rejection rate are equal. A lower EER generally indicates better verification performance.
The findings suggest that longer recordings usually helped the systems distinguish languages more reliably.
This is an important lesson in speech technology: the amount of available speech can influence how confidently a computer identifies a language.
6. What Similarities Did the Research Discover?
The researchers analysed the 30-second recordings in greater detail and identified several notable patterns.
They considered two languages similar under their experimental framework when the false-acceptance percentage reached at least 25% in the relevant comparison in three or more of the four verification systems.
The following examples were reported in the paper.
Tamil and Malayalam
The researchers found strong mutual confusion between Tamil and Malayalam.
When Malayalam was tested against the Tamil-target systems, the reported false-acceptance percentages ranged from 75% to 90% across three of the systems. In the reverse comparison, Tamil showed percentages ranging from 55% to 75%.
These results indicated substantial acoustic similarity under the study's method.
This finding is understandable in a linguistic context because Tamil and Malayalam belong to the Dravidian language family and are spoken in neighbouring regions.
Their histories include long periods of geographical and cultural contact.
However, the computer experiment measured patterns in speech recordings. It did not establish the complete historical relationship between the two languages; that requires linguistic and historical evidence as well.
Gujarati and Hindi
The study also identified similarity between Gujarati and Hindi.
Gujarati showed false-acceptance percentages of approximately 25%–50% against Hindi, while Hindi showed approximately 30%–55% against Gujarati in the reported analysis.
These results suggest that the systems had difficulty distinguishing certain speech samples from these two languages.
Gujarati and Hindi belong to the Indo-Aryan branch of the Indo-European language family. Their speakers have also had opportunities for contact through geographical proximity, migration and cultural exchange.
The research therefore identified a pattern consistent with both linguistic relatedness and regional contact.
Gujarati and Marathi
Another reported similarity was between Gujarati and Marathi.
Gujarati showed false-acceptance percentages of approximately 25%–75% against Marathi, while Marathi showed approximately 50%–70% against Gujarati in the three relevant systems.
This indicates that the two languages shared enough acoustic characteristics in the selected recordings to confuse the verification systems relatively often.
Gujarati and Marathi are Indo-Aryan languages spoken in neighbouring western Indian regions, providing a linguistic and geographical context for the observation.
Hindi and Marathi
The researchers also found similarity between Hindi and Marathi.
Hindi showed false-acceptance percentages of approximately 30%–75% against Marathi, while Marathi showed approximately 30%–50% against Hindi.
Once again, the result reflected the behaviour of the speech-verification systems rather than a direct measurement of how much vocabulary or grammar the languages share.
English and Kashmiri: An unexpected result
One particularly interesting observation involved English and Kashmiri.
The researchers reported false-acceptance percentages of approximately 65%–90% when English recordings were tested against Kashmiri-target systems. In the reverse direction, the reported range was approximately 60%–75%.
The authors discussed possible explanations involving historical and social contact, tourism and the use of English as a language of communication in Kashmir.
This result needs careful interpretation.
English belongs to the Germanic branch of the Indo-European language family, whereas Kashmiri is an Indo-Aryan language. Their relationship is not equivalent to the relationship between Tamil and Malayalam.
Moreover, the study measured acoustic confusion, not proof that English and Kashmiri share a particularly close linguistic ancestry.
The finding is best understood as an unexpected result that the authors attempted to explain using historical and social context. It would require further research to establish precisely why the systems confused these recordings.
7. Can a Computer Discover Similarities That Humans Have Missed?
This is one of the most interesting possibilities raised by the research.
Traditional linguistic research often examines vocabulary, grammar, pronunciation, historical documents and reconstructed ancestral languages.
Computer-based analysis offers another approach: it can process many speech samples and systematically measure recurring patterns.
Suppose researchers suspect that two languages share a feature but are unsure how consistently that feature appears in everyday speech. A computational system could analyse recordings and identify patterns worthy of further investigation.
However, a computer's finding should be treated as a research lead, not automatically as a linguistic discovery.
Speech recordings can be influenced by several factors, including:
The speaker's regional accent.
Whether the speaker uses a formal or conversational style.
The speed and rhythm of speech.
The recording quality and background noise.
Words borrowed from another language.
The selection of speakers and topics in the dataset.
A computer may confuse languages because of shared pronunciation patterns, similar accents or characteristics of the selected recordings. These possibilities must be investigated before drawing broader conclusions.
The value of the method is that it can help researchers identify patterns for further linguistic examination.
8. What Did the Research Teach Us About Language Families and Geography?
The study supports an important idea: similarities between languages can arise through more than one historical process.
First, languages can share an ancestral origin. Languages descended from a common ancestor may retain similar sounds and linguistic structures even after developing independently.
Second, geographical proximity can encourage language contact. People who live near one another often interact through trade, migration, work, education and cultural exchange.
Third, social and historical circumstances can create unexpected patterns. Languages that do not share a particularly close relationship may still exhibit similarities in certain speech features.
These explanations are not interchangeable. Shared ancestry, borrowed vocabulary, regional pronunciation and similarities in acoustic patterns are different phenomena.
The research is particularly interesting because it demonstrates how speech analysis can contribute another source of evidence to the study of language relationships.
9. What Were the Limitations of the Study?
The paper provides useful findings, but its results must be understood within the limits of its experimental design.
Only 16 languages were tested. India has a much larger linguistic landscape, so the findings cannot represent every Indian language or dialect.
The recordings came from news broadcasts. Newsreaders may speak differently from people in everyday conversations. The results might change if the database included informal speech from a wider variety of speakers and regions.
The system analysed sound rather than meaning. It did not establish similarity by comparing the meanings of thousands of words, grammatical rules or historical language records.
False acceptance is an indirect measure. A high rate of confusion can indicate acoustic similarity, but it does not independently prove shared ancestry or close linguistic relationships.
Some results were asymmetric. A system trained for Language A might frequently accept Language B, while the system trained for Language B might not show the same behaviour in reverse. The authors discussed such inconsistencies and examined different false-acceptance thresholds.
These limitations do not make the research unimportant. They clarify what the experiment demonstrates and where additional evidence is necessary.
10. Why Is This Research Relevant Today?
Speech-processing technology has applications in many areas, including multilingual communication, speech recognition, language identification and the development of digital tools for underrepresented languages.
Research of this kind can contribute to systems that identify the language being spoken, particularly in multilingual environments where several languages may share certain sound patterns.
It may also help researchers organise large collections of recordings, compare regional speech patterns and identify cases that deserve closer linguistic study.
For India, this area of research has particular relevance because many languages and dialects have limited digital resources compared with widely used global languages.
Computational methods could help document and analyse speech varieties that might otherwise receive less attention.
Nevertheless, a system developed using 16 languages and news recordings would need further testing before being used as a general-purpose language identification tool.
Additional Findings: AI Reveals More Unexpected Similarities Between Indian Languages
When the researchers increased the false-acceptance threshold from 15% to 20%, their analysis revealed additional similarities among Indian languages. The results showed that the relationship between languages was not limited to languages belonging to the same linguistic family or spoken in neighbouring regions. Some findings were consistent with existing linguistic and historical knowledge, while others raised questions that could be explored through further research.
More Evidence of Similarity Between Tamil and Malayalam
The study reported that Malayalam recordings showed an 85%–100% false-acceptance percentage against Tamil-target systems, while Tamil recordings showed 45%–75% against Malayalam-target systems. This indicates substantial acoustic similarity between the two languages under the researchers’ method. Both belong to the Dravidian language family and have developed through long periods of interaction in southern India.
English and Kashmiri: An Unexpected Connection
One of the most striking findings involved English and Kashmiri. English showed a 75%–95% false-acceptance percentage against Kashmiri, while Kashmiri showed 70%–90% against English. The researchers discussed possible explanations involving tourism, social interaction and the use of English for communication. However, these results measure similarities in recorded speech, not proof that the two languages have a particularly close historical relationship.
Gujarati, Hindi and Marathi: A Cluster of Similarities
The researchers found similarities among Gujarati, Hindi and Marathi. Gujarati showed a 35%–50% false-acceptance percentage against Hindi, while Hindi showed 35%–65% against Gujarati. Gujarati and Marathi also showed substantial mutual confusion, with reported ranges of 25%–75% and 50%–75%, respectively. Similarly, Hindi showed 45%–85% against Marathi, while Marathi showed 50%–75% against Hindi.
These findings are consistent with the fact that the three languages belong to the Indo-Aryan language group and have developed in regions with historical and geographical connections. The study demonstrates how computational analysis can identify recurring patterns across multiple language pairs.
Konkani and Gujarati: A Relationship Supported by Linguistic Research
The researchers reported a 25%–45% false-acceptance percentage for Gujarati against Konkani and 45%–55% for Konkani against Gujarati. They connected this finding to earlier linguistic literature that discusses Konkani's position within the Indo-Aryan language group and its relationship with Gujarati.
The result illustrates how computer-based observations can be compared with existing linguistic classifications rather than interpreted in isolation.
Assamese and Odia: Similarities Across Eastern India
Assamese showed a 35%–65% false-acceptance percentage against Odia, while Odia showed 30%–70% against Assamese. The researchers considered this finding consistent with existing discussions of similarities between the two languages.
This example highlights the value of combining computational evidence with established linguistic research when examining relationships between regional languages.
English and Konkani: The Influence of Historical Contact
Another notable finding concerned English and Konkani. English showed a 25%–30% false-acceptance percentage against Konkani, while Konkani showed a much wider range of 25%–100% against English.
The authors suggested that Goa's long history of overseas trade, colonial contact and tourism could help explain interactions between Konkani and English. These historical circumstances provide a possible context for language contact, although the experiment itself does not establish the cause of the observed acoustic similarity.
Hindi and Urdu: Closely Connected Speech Traditions
The study also identified similarity between Hindi and Urdu. Hindi showed a 25%–40% false-acceptance percentage against Urdu, while Urdu showed 30%–60% against Hindi.
This result is particularly interesting because the standard spoken forms of Hindi and Urdu share substantial vocabulary and grammatical structures. Their differences are more pronounced in aspects such as formal vocabulary choices and writing systems. The computer's difficulty in distinguishing some recordings is therefore consistent with similarities in their spoken forms.
Kannada and Marathi: Similarities Between Neighbouring Regions
Kannada showed a 30%–45% false-acceptance percentage against Marathi. In the reverse direction, Marathi showed a 25% false-acceptance percentage against Kannada in three of the four verification systems.
The researchers linked this observation to the geographical proximity of Karnataka and Maharashtra and the historical interaction between their populations. The result suggests that the speech characteristics captured by the systems overlapped in some recordings, although it does not establish the precise linguistic causes.
Malayalam and Sanskrit: A Historical Vocabulary Connection
Malayalam showed a 25%–50% false-acceptance percentage against Sanskrit, while Sanskrit showed 40%–70% against Malayalam.
The authors referred to the substantial influence of Sanskrit on Malayalam vocabulary as a possible explanation for the observed similarity. Malayalam has a long history of contact with Sanskrit, particularly through literary, religious and scholarly traditions.
Nevertheless, vocabulary borrowing and acoustic similarity are different things. The study did not directly measure the proportion of Sanskrit-derived words in Malayalam; it used spoken-speech verification to identify language confusion.
Tamil and Telugu: Similarities Within the Dravidian Family
Tamil showed a 30%–55% false-acceptance percentage against Telugu, while Telugu showed 30%–40% against Tamil. Both languages belong to the Dravidian family and have a long history of contact in southern India.
The findings are consistent with a broader pattern in the study: languages with related linguistic histories or geographical connections often exhibited overlapping speech characteristics.
Bengali and Gujarati: A Question for Further Research
One particularly intriguing observation involved Bengali and Gujarati. Bengali showed a 30%–70% false-acceptance percentage against Gujarati, while Gujarati showed 25%–45% against Bengali.
The researchers noted that this relationship could not be readily explained using the linguistic information they consulted at the time. They suggested that it might represent a pattern worthy of further investigation.
This should not be interpreted as proof of a previously unknown historical relationship. The result could reflect acoustic similarities, characteristics of the recordings or other factors not captured by the experiment. Independent linguistic analysis would be needed to determine the explanation.
What Makes These Results Interesting?
The study shows how changing an experimental threshold can reveal additional patterns in a dataset. At the 20% false-acceptance threshold, the researchers identified more potentially similar language pairs than at 15%.
However, a higher threshold also allows more recordings to be included among the accepted samples, so additional findings must be interpreted carefully. The researchers compared results from four verification systems and used linguistic and historical context to evaluate their observations.
The broader contribution of this research is its demonstration that spoken-language analysis can serve as a supplementary tool for investigating language relationships. It does not replace historical linguistics, but it can help identify patterns that researchers may wish to examine using vocabulary comparisons, grammatical evidence, historical documents and larger speech datasets.
Research Reference
DOI: https://doi.org/10.1155/2015/325703
