Finds sound-alike spellings of Bangla and English words, with a match percentage. One Kotlin Multiplatform library for Android (Views and Compose), the JVM and desktop, iOS and the other Apple platforms, the web, Linux and Windows. The search runs off the calling thread, so the UI never waits for it.

// Bangla phonetics
val bangla = BanglaPhonetics.instance
val result = bangla.generate("মাসুদ", minScore = 90.0)
// মাসুদ 100%, মাসূদ 97%, মাশুদ 95%, মাষুদ 95%, মাশূদ 92.15%, মাষূদ 92.15%

// English phonetics
val english = EnglishPhonetics.instance
val names = english.generate("Philip", minScore = 90.0)
// Philip 100%, Filip 95%, Phillip 95%, Philipp 95%, Fillip 90.25%, ...

generate is a suspend function. Every call returns a result object with the value and an error; nothing is thrown.

Add the dependency to the module that needs it. In a Kotlin Multiplatform project that is commonMain:

kotlin {
    sourceSets {
        commonMain.dependencies {
            implementation("io.github.zakaria5729:phonetics:0.1.1")
            // for Compose
            implementation("io.github.zakaria5729:phonetics-compose:0.1.1")
        }
    }
}

An Android or JVM only project uses the same coordinates:

dependencies {
    implementation("io.github.zakaria5729:phonetics:0.1.1")
}

Groovy DSL (build.gradle) and Maven (pom.xml) for a JVM project:

implementation 'io.github.zakaria5729:phonetics-jvm:0.1.1'

<dependency>
    <groupId>io.github.zakaria5729</groupId>
    <artifactId>phonetics-jvm</artifactId>
    <version>0.1.1</version>
</dependency>

Download from Maven Central:

The jar, sources and Javadoc files for every target are also in the repository folder.

The Android library needs minSdk 21. phonetics-compose needs a Compose Multiplatform version that supports your targets (it has no Intel iOS or macOS build, because Compose has none).

The artifacts are on Maven Central, so mavenCentral() is all you need in your repositories. To try local changes, run ./gradlew publishToMavenLocal in this repository and add mavenLocal() as well.

val phonetics = BanglaPhonetics.instance

val matches = phonetics.generate("আসাদ", minScore = 90.0, maxResults = 20)
matches.error?.let { println(it.message) }
matches.matches.forEach { println("${it.word} ${it.score}%") }

val score = phonetics.similarity("আসাদ", "আছাদ").score          // 85.0
val same = phonetics.matches("আসাদ", "আশাদ", minScore = 90.0).isMatch
val ranked = phonetics.rank("আসাদ", listOf("আছাদ", "রহিম", "আশাদ"), limit = 2)

Defaults are minScore 70 and maxResults 50. Each match has the word, the score from 0 to 100 and the changes that turn the input into the word.

Callbacks work without coroutines. They run on the main thread on Android and Apple platforms, and the call returns a token that cancels the request:

val bangla = BanglaPhonetics()

input.addTextChangedListener { text ->
    bangla.generateAsync(text.toString(), maxResults = 20) { result ->
        results.text = result.matches.joinToString("\n") { "${it.word} ${it.score}%" }
    }
}

override fun onDestroy() {
    bangla.dispose()
    super.onDestroy()
}

sample-android-xml is a full example. In a view model, call the suspend functions from viewModelScope or lifecycleScope instead.

For Java, use the …Async functions with a lambda that returns Unit.INSTANCE. For Swift, the suspend functions become async functions and the callbacks become closures:

let result = try await BanglaPhonetics.companion.instance.generate(input: "আসাদ")
print(result.matches.map { $0.word })

BanglaPhonetics.companion.instance.generateAsync(input: "আসাদ") { result in
    print(result.matches.count)
}

Build the XCFramework on a Mac with Xcode: ./gradlew :phonetics:assemblePhoneticsXCFramework. The Swift snippets above have not been compiled yet, because this was written on a machine without Xcode.

@Composable
fun Names() {
    val phonetics = rememberBanglaPhonetics()
    var input by remember { mutableStateOf("আসাদ") }
    val result by rememberPhoneticMatches(phonetics, input, maxResults = 30, debounceMillis = 120)

    Column {
        OutlinedTextField(value = input, onValueChange = { input = it })
        LazyColumn {
            items(result?.matches.orEmpty()) { match -> Text("${match.word} ${match.score}%") }
        }
    }
}

rememberPhonetics, rememberBanglaPhonetics and rememberEnglishPhonetics dispose the instance when it leaves the composition. rememberPhoneticMatches, rememberPhoneticSimilarity and rememberPhoneticRanking return null until the first result, keep the previous result while a new one is computed, and cancel a running request when the input changes. It is the same code on Android, desktop, iOS and the web; sample-compose shows it.

PhoneticLanguages.bangla and PhoneticLanguages.english ship with the library. A language is data: a code, a name, the rules, and how its text is cleaned up before use.

val language = PhoneticLanguage(
    code = "xx",
    name = "Test",
    foldCase = true,
    canonicalSequences = mapOf("ß" to "ss"),
    marks = setOf(0x2D),
    rules = listOf(
        PhoneticRule(listOf("q", "x"), 0.9),
        PhoneticRule(listOf("ee", "i"), 0.8, position = RulePosition.WordEnd),
    ),
)
val phonetics = Phonetics(language)

Bangla rules treat ড় and ঢ় written with a nukta the same as the single letters. English folds case, removes accents and gives the answer back in the case of the input (JOHN gives JOHN, John Smith gives John Smith).

A PhoneticRule lists spellings that sound alike and how alike (above 0 and below 1). position limits it to the start, the end or the inside of a word. scoreOnly rules count when comparing words but are never used to build new spellings. Extend the defaults with PhoneticLanguages.bangla.defaultRuleSet.extend(listOf(...)) and pass the result as rules.

Nothing is thrown. Bad arguments, a broken language or rule set, a cancelled request and a disposed instance all come back in error:

Constructors keep their problems in an error property too, so a bad setting is reported on the first call instead of crashing at construction.

One thing is not turned into a result: if the coroutine that called a suspend function is cancelled, Kotlin's CancellationException propagates as usual, and the work stops.

val token = PhoneticCancelToken()
val call = scope.async { phonetics.rank(query, candidates, cancelToken = token) }
token.cancel()
call.await().error   // PhoneticCancelledException

phonetics.cancelAll() cancels everything that is pending, and dispose() cancels it and stops the instance.

PhoneticExecutionMode decides what a new request does while one is running:

Pass it to the constructor or change it later with phonetics.setExecutionMode(...), also on the shared BanglaPhonetics.instance.

PhoneticEngine is the same code without the background execution. Use it for short work, in tests, or from your own dispatcher:

val engine = PhoneticEngine(PhoneticLanguages.english)
val result = engine.generate("Catherine")

The number of variants grows quickly as minScore falls, so raising it is the strongest way to save memory. A call returns at most 100,000 results and about 20 million output characters, whatever you ask for.

rank keeps only limit matches, so use a limit with big candidate lists.

n is the length of the input, R the number of results, m the length of the other word and C the number of candidates.

generate is a best-first search over letter swaps backed by a heap. It always expands the highest-scoring swaps first and stops as soon as maxResults are collected or the scores fall below minScore. rank stops scoring a candidate as soon as it cannot reach the cut-off.

Every swap has a similarity, for example স→শ is 0.95 and স→ছ is 0.85. A variant's score is the product of its swaps, so two স→শ swaps give 90.25%. The default rules are banglaRules and englishRules.

Scores are rounded to two decimals. When a score falls exactly between two roundings, the last digit can differ by 0.01 from the Dart package, because the platforms' exp and ln differ in the last bit.

90% accuracy gain, measured by testing against public datasets.

Datasets:

• Bangla
  • https://github.com/google/language-resources/tree/master/bn/data
  • https://github.com/google-research-datasets/dakshina
  • https://huggingface.co/datasets/ai4bharat/Aksharantar
• English
  • https://github.com/cmusphinx/cmudict
  • https://llds.phon.ox.ac.uk/llds/xmlui/handle/20.500.14106/0643
  • https://github.com/mhagiwara/github-typo-corpus

This package is also available for Flutter: pub.dev | GitHub

./gradlew :phonetics:jvmTest
./gradlew :phonetics:jsNodeTest :phonetics:wasmJsNodeTest
./gradlew :phonetics:jsBrowserTest :phonetics:wasmJsBrowserTest   # needs Chrome
./gradlew :phonetics:testAndroidHostTest
./gradlew :phonetics-compose:jvmTest
./gradlew :sample-compose:run                                    # desktop sample
./gradlew :sample-compose:wasmJsBrowserDevelopmentRun            # web sample
./gradlew :sample-android-xml:installDebug

phonetics/src/commonTest/.../GoldenData.kt holds results from the Dart package. To refresh it, run dart run bin/generate.dart <path> in tools/golden.

MIT, see LICENSE.

Built and maintained by Zakaria Hossain, known online as zakaria5729.

LinkedIn | GitHub | Stack Overflow | Medium | Substack | Twitter | Facebook

LinkedIn | GitHub | Stack Overflow | Medium | Substack | Twitter | Facebook