RunWhisper.cs

using System.Collections;
using System.Collections.Generic;
using UnityEngine;
using Unity.Sentis;
using System.IO;
using Newtonsoft.Json;
using System.Text;

/*
 *              Whisper Inference Code
 *              ======================
 *  
 *  Put this script on the Main Camera
 *  
 *  In Assets/StreamingAssets put:
 *  
 *  AudioDecoder_Tiny.sentis
 *  AudioEncoder_Tiny.sentis
 *  LogMelSepctro.sentis
 *  vocab.json
 * 
 *  Drag a 30s 16khz mono uncompressed audioclip into the audioClip field. 
 * 
 *  Install package com.unity.nuget.newtonsoft-json from packagemanger
 *  Install package com.unity.sentis
 * 
 */


public class RunWhisper : MonoBehaviour
{
    IWorker decoderEngine, encoderEngine, spectroEngine;

    const BackendType backend = BackendType.GPUCompute;

    // Link your audioclip here. Format must be 16Hz mono non-compressed.
    public AudioClip audioClip;

    const int maxTokens = 100;

    //Special tokens
    const int END_OF_TEXT = 50257;
    const int START_OF_TRANSCRIPT = 50258;
    const int ENGLISH = 50259;
    const int TRANSCRIBE = 50359;
    const int START_TIME = 50364;

    Ops ops;
    ITensorAllocator allocator;

    int numSamples;
    float[] data;
    string[] tokens;

    int currentToken = 0;
    int[] outputTokens = new int[maxTokens];

    // Used for special character decoding
    int[] shiftDownDict = new int[256];

    TensorFloat encodedAudio;

    bool transcribe = false;
    string outputString = "";

    void Start()
    {
        allocator = new TensorCachingAllocator();
        ops = WorkerFactory.CreateOps(backend, allocator);

        SetupCharacterShifts();

        GetTokens();

        Model decoder = ModelLoader.Load(Application.streamingAssetsPath + "/AudioDecoder_Tiny.sentis");
        Model encoder = ModelLoader.Load(Application.streamingAssetsPath + "/AudioEncoder_Tiny.sentis");
        Model spectro = ModelLoader.Load(Application.streamingAssetsPath + "/LogMelSepctro.sentis");

        decoderEngine = WorkerFactory.CreateWorker(backend, decoder);
        encoderEngine = WorkerFactory.CreateWorker(backend, encoder);
        spectroEngine = WorkerFactory.CreateWorker(backend, spectro);

        outputTokens[0] = START_OF_TRANSCRIPT;
        outputTokens[1] = ENGLISH;
        outputTokens[2] = TRANSCRIBE;
        outputTokens[3] = START_TIME;
        currentToken = 3;

        LoadAudio();
        EncodeAudio();
        transcribe = true;
    }

    void LoadAudio()
    {
        if(audioClip.frequency != 16000)
        {
            Debug.Log($"The audio clip should have frequency 16kHz. It has frequency {audioClip.frequency / 1000f}kHz");
        }

        numSamples = audioClip.samples;
        data = new float[numSamples];
        audioClip.GetData(data, 0);
    }


    void GetTokens()
    {
        var jsonText = File.ReadAllText(Application.streamingAssetsPath + "/vocab.json");
        var vocab = Newtonsoft.Json.JsonConvert.DeserializeObject<Dictionary<string, int>>(jsonText);
        tokens = new string[vocab.Count];
        foreach(var item in vocab)
        {
            tokens[item.Value] = item.Key;
        }
    }

    void EncodeAudio()
    {
        var input = new TensorFloat(new TensorShape(1, numSamples), data);

        int maxSamples = 30 * 16000;
        if (numSamples > maxSamples)
        {
            Debug.Log("The AudioClip is too long.");
            return;
        }

        // Pad out to 30 seconds at 16khz if necessary
        var input30seconds = ops.Pad(input, new int[] { 0, 0, 0, 30 * 16000 - numSamples });

        spectroEngine.Execute(input30seconds);
        var spectroOutput = spectroEngine.PeekOutput() as TensorFloat;

        encoderEngine.Execute(spectroOutput);
        encodedAudio = encoderEngine.PeekOutput() as TensorFloat;
    }


    // Update is called once per frame
    void Update()
    {
        if (transcribe && currentToken < outputTokens.Length - 1)
        {
            var tokensSoFar = new TensorInt(new TensorShape(1, outputTokens.Length), outputTokens);

            var inputs = new Dictionary<string, Tensor>
            {
                {"encoded_audio",encodedAudio },
                {"tokens" , tokensSoFar }
            };

            decoderEngine.Execute(inputs);
            var tokensOut = decoderEngine.PeekOutput() as TensorFloat;

            var tokensPredictions = ops.ArgMax(tokensOut, 2, false);
            tokensPredictions.MakeReadable();

            int ID = tokensPredictions[currentToken];

            currentToken++;
            outputTokens[currentToken] = ID;

            if (ID == END_OF_TEXT)
            {
                transcribe = false;
            }
            else if (ID >= tokens.Length) outputString += $"(time={(ID - START_TIME) * 0.02f})";
            else outputString += GetUnicodeText(tokens[ID]);

            Debug.Log(outputString);
        }
    }

    // Translates encoded special characters to Unicode
    string GetUnicodeText(string text)
    {
        var bytes = Encoding.GetEncoding("ISO-8859-1").GetBytes(ShiftCharacterDown(text));
        return Encoding.UTF8.GetString(bytes);
    }

    string ShiftCharacterDown(string text)
    {
        string outText = "";
        foreach (char letter in text)
        {
            outText += ((int)letter <= 256) ? letter :
                (char)shiftDownDict[(int)(letter - 256)];
        }
        return outText;
    }

    void SetupCharacterShifts()
    {
        for (int i = 0, n = 0; i < 256; i++)
        {
            if (IsWhiteSpace((char)i)) shiftDownDict[n++] = i;
        }
    }

    bool IsWhiteSpace(char c)
    {
        return !(('!' <= c && c <= '~') || ('�' <= c && c <= '�') || ('�' <= c && c <= '�'));
    }

    private void OnDestroy()
    {
        decoderEngine?.Dispose();
        encoderEngine?.Dispose();
        spectroEngine?.Dispose();
        ops?.Dispose();
    }
}