CIIL Anuvani-LID

bd48fd1 21 days ago

3.66 kB

// Copyright (c) 2026, CIIL Anuvani-LID. Apache-2.0.
//
// Live language identification from the microphone: AudioRecord (16 kHz mono float) -> sherpa-onnx's Silero VAD ->
// each finished stretch of speech is classified on a worker thread (so a slow classification never stalls the
// recording) -> onSegment(start, end, scores) on that worker thread. Needs android.permission.RECORD_AUDIO.
// A starting point: copy it into your app and change the package (docs/03-android.md, "Live microphone").
package com.example.anuvanilid

import android.annotation.SuppressLint
import android.media.AudioFormat
import android.media.AudioRecord
import android.media.MediaRecorder
import com.k2fsa.sherpa.onnx.SileroVadModelConfig
import com.k2fsa.sherpa.onnx.Vad
import com.k2fsa.sherpa.onnx.VadModelConfig
import java.util.concurrent.Executors
import java.util.concurrent.TimeUnit

class LiveLanguageId(
    private val identifier: LanguageIdentifier,
    vadModel: String,
    /** Called once per stretch of speech (seconds from start(); the most likely languages first). */
    private val onSegment: (startS: Double, endS: Double, scores: List<LanguageScore>) -> Unit,
    /** Longest stretch passed to the model in one piece (seconds). */
    maxSpeechSeconds: Float = 15f,
) : AutoCloseable {
    private val sampleRate = 16000
    private val vad = Vad(
        config = VadModelConfig(
            sileroVadModelConfig = SileroVadModelConfig(
                model = vadModel,
                threshold = 0.5f,
                minSilenceDuration = 0.3f,
                minSpeechDuration = 0.25f,
                windowSize = 512,
                maxSpeechDuration = maxSpeechSeconds,
            ),
            sampleRate = sampleRate,
            numThreads = 1,
        ),
    )
    private val worker = Executors.newSingleThreadExecutor()

    @Volatile
    private var running = false
    private var thread: Thread? = null

    @SuppressLint("MissingPermission")
    fun start() {
        check(!running) { "already started" }
        running = true
        thread = Thread {
            val minBuf = AudioRecord.getMinBufferSize(sampleRate, AudioFormat.CHANNEL_IN_MONO, AudioFormat.ENCODING_PCM_FLOAT)
            val rec = AudioRecord(
                MediaRecorder.AudioSource.VOICE_RECOGNITION, sampleRate, AudioFormat.CHANNEL_IN_MONO,
                AudioFormat.ENCODING_PCM_FLOAT, maxOf(minBuf, 4 * sampleRate),
            )
            val buf = FloatArray(512)
            rec.startRecording()
            while (running) {
                val n = rec.read(buf, 0, buf.size, AudioRecord.READ_BLOCKING)
                if (n <= 0) continue
                vad.acceptWaveform(if (n == buf.size) buf.copyOf() else buf.copyOf(n))
                drain()
            }
            rec.stop()
            rec.release()
            vad.flush()
            drain()
        }.apply { name = "anuvani-lid-mic"; start() }
    }

    /** Stop recording; segments already queued are still classified. */
    fun stop() {
        running = false
        thread?.join()
        thread = null
    }

    override fun close() {
        stop()
        worker.shutdown()
        worker.awaitTermination(60, TimeUnit.SECONDS)
        vad.release()
    }

    private fun drain() {
        while (!vad.empty()) {
            val seg = vad.front()
            vad.pop()
            val start = seg.start / sampleRate.toDouble()
            val end = (seg.start + seg.samples.size) / sampleRate.toDouble()
            val samples = seg.samples
            worker.execute { onSegment(start, end, identifier.classify(samples, sampleRate, topK = 3)) }
        }
    }
}

	// Copyright (c) 2026, CIIL Anuvani-LID. Apache-2.0.
	//
	// Live language identification from the microphone: AudioRecord (16 kHz mono float) -> sherpa-onnx's Silero VAD ->
	// each finished stretch of speech is classified on a worker thread (so a slow classification never stalls the
	// recording) -> onSegment(start, end, scores) on that worker thread. Needs android.permission.RECORD_AUDIO.
	// A starting point: copy it into your app and change the package (docs/03-android.md, "Live microphone").
	package com.example.anuvanilid

	import android.annotation.SuppressLint
	import android.media.AudioFormat
	import android.media.AudioRecord
	import android.media.MediaRecorder
	import com.k2fsa.sherpa.onnx.SileroVadModelConfig
	import com.k2fsa.sherpa.onnx.Vad
	import com.k2fsa.sherpa.onnx.VadModelConfig
	import java.util.concurrent.Executors
	import java.util.concurrent.TimeUnit

	class LiveLanguageId(
	private val identifier: LanguageIdentifier,
	vadModel: String,
	/** Called once per stretch of speech (seconds from start(); the most likely languages first). */
	private val onSegment: (startS: Double, endS: Double, scores: List<LanguageScore>) -> Unit,
	/** Longest stretch passed to the model in one piece (seconds). */
	maxSpeechSeconds: Float = 15f,
	) : AutoCloseable {
	private val sampleRate = 16000
	private val vad = Vad(
	config = VadModelConfig(
	sileroVadModelConfig = SileroVadModelConfig(
	model = vadModel,
	threshold = 0.5f,
	minSilenceDuration = 0.3f,
	minSpeechDuration = 0.25f,
	windowSize = 512,
	maxSpeechDuration = maxSpeechSeconds,
	),
	sampleRate = sampleRate,
	numThreads = 1,
	),
	)
	private val worker = Executors.newSingleThreadExecutor()

	@Volatile
	private var running = false
	private var thread: Thread? = null

	@SuppressLint("MissingPermission")
	fun start() {
	check(!running) { "already started" }
	running = true
	thread = Thread {
	val minBuf = AudioRecord.getMinBufferSize(sampleRate, AudioFormat.CHANNEL_IN_MONO, AudioFormat.ENCODING_PCM_FLOAT)
	val rec = AudioRecord(
	MediaRecorder.AudioSource.VOICE_RECOGNITION, sampleRate, AudioFormat.CHANNEL_IN_MONO,
	AudioFormat.ENCODING_PCM_FLOAT, maxOf(minBuf, 4 * sampleRate),
	)
	val buf = FloatArray(512)
	rec.startRecording()
	while (running) {
	val n = rec.read(buf, 0, buf.size, AudioRecord.READ_BLOCKING)
	if (n <= 0) continue
	vad.acceptWaveform(if (n == buf.size) buf.copyOf() else buf.copyOf(n))
	drain()
	}
	rec.stop()
	rec.release()
	vad.flush()
	drain()
	}.apply { name = "anuvani-lid-mic"; start() }
	}

	/** Stop recording; segments already queued are still classified. */
	fun stop() {
	running = false
	thread?.join()
	thread = null
	}

	override fun close() {
	stop()
	worker.shutdown()
	worker.awaitTermination(60, TimeUnit.SECONDS)
	vad.release()
	}

	private fun drain() {
	while (!vad.empty()) {
	val seg = vad.front()
	vad.pop()
	val start = seg.start / sampleRate.toDouble()
	val end = (seg.start + seg.samples.size) / sampleRate.toDouble()
	val samples = seg.samples
	worker.execute { onSegment(start, end, identifier.classify(samples, sampleRate, topK = 3)) }
	}
	}
	}