c# - 如何使用 Watson Unity SDK 制作语音转文本自定义模型？

Question

我在 Unity 中使用 Watson Assistant、语音转文本和文本转语音制作了一个应用程序，用户可以在其中说出不同的城市以查找所述城市之间的可用机票。对话和互动效果很好，但有时我会遇到用户说出某些城市时无法识别的问题。例如柏林，有时它理解柏林，有时它理解燃烧。巴黎、伦敦和 Jakarta 等其他城市也是如此。

因此，城市名称的检测并不总是像我希望的那样准确。但我在一些帖子中看到，您可以制作自己的自定义模型来改进对这些词的检测。但我不知道如何设置、制作自己的自定义模型以及如何将这些城市添加到模型中并进行训练。是否可以在 Unity C# 脚本中执行此操作，我将如何开始？有一些我可以看的 C# 示例吗？任何帮助将不胜感激。

这些是我找到的一些链接和信息，但不知道如何在 C# 中实现它以及我自己关于提高城市检测准确性的目的。

DwAnswers1 DwAnswers2 StackOverflow IBM clouds docs Medium cURL tutorial

这是我在 Watson API 和 Unity 之间进行交互的 C# 脚本。我想我也必须在这里添加自定义模型，但我不知道我是否也应该在其中创建自定义模型，或者它是否需要在单独的脚本中。

using System.Collections;
using System.Collections.Generic;
using UnityEngine;
using IBM.Watson.DeveloperCloud.Services.TextToSpeech.v1;
using IBM.Watson.DeveloperCloud.Services.Conversation.v1;
using IBM.Watson.DeveloperCloud.Services.ToneAnalyzer.v3;
using IBM.Watson.DeveloperCloud.Services.SpeechToText.v1;
using IBM.Watson.DeveloperCloud.Logging;
using IBM.Watson.DeveloperCloud.Utilities;
using IBM.Watson.DeveloperCloud.Connection;
using IBM.Watson.DeveloperCloud.DataTypes;
using MiniJSON;
using UnityEngine.UI;
using FullSerializer;

public class WatsonAgent : MonoBehaviour
{

public string literalEntityCity;
public string destinationCity;
public string departureCity;

public string dateBegin;
public string dateEnd;

public WeatherJSON weather;
public GameObject FlightInfo;

[SerializeField]
private fsSerializer _serializer = new fsSerializer();

[System.Serializable]
public class CredentialInformation
{
    public string username, password, url;
}

[System.Serializable]
public class Services
{
    public CredentialInformation
        textToSpeech,
        conversation,
        speechToText;
}

[Header("Credentials")]
[Space]
public Services
    serviceCredentials;

[Space]
[Header("Agent voice settings")]
[Space]
public AudioSource
    voiceSource;

public VoiceType
    voiceType;

[Space]
[Header("Conversation settings")]
[Space]
public string
    workspaceId;

[Space]
[Header("Feedback fields")]
[Space]
public Text
    speechToTextField;
public Text
    conversationInputField;
public Text
    conversationOutputField;

public string
    saying;

// services
SpeechToText
    speechToText;

private int
    recordingRoutine = 0,
    recordingBufferSize = 1,
    recordingHZ = 22050;

private string
    microphoneID = null;

private AudioClip
    recording = null;

TextToSpeech
    textToSpeech;

Conversation
    conversation;

private Dictionary<string, object>
    conversationContext = null;

private void Start()
{
    PrepareCredentials();
    Initialize();
}

void PrepareCredentials()
{
    speechToText = new SpeechToText(GetCredentials(serviceCredentials.speechToText));
    textToSpeech = new TextToSpeech(GetCredentials(serviceCredentials.textToSpeech));
    conversation = new Conversation(GetCredentials(serviceCredentials.conversation));
}

Credentials GetCredentials(CredentialInformation credentialInformation)
{
    return new Credentials(credentialInformation.username, credentialInformation.password, credentialInformation.url);
}

void Initialize()
{
    conversation.VersionDate = "2017-05-26";
    Active = true;
    StartRecording();
}

// speech to text
public bool Active
{
    get { return speechToText.IsListening; }
    set
    {
        if (value && !speechToText.IsListening)
        {
            speechToText.DetectSilence = true;
            speechToText.EnableWordConfidence = true;
            speechToText.EnableTimestamps = true;
            speechToText.SilenceThreshold = 0.01f;
            speechToText.MaxAlternatives = 0;
            speechToText.EnableInterimResults = true;
            speechToText.OnError = OnSpeechError;
            speechToText.InactivityTimeout = -1;
            speechToText.ProfanityFilter = false;
            speechToText.SmartFormatting = true;
            speechToText.SpeakerLabels = false;
            speechToText.WordAlternativesThreshold = null;
            speechToText.StartListening(OnSpeechRecognize);
            //speechToText.CustomizationId = "customID";    // I guess i have to add the custom training model here with the customID
            //speechToText.CustomizationWeight(0.2);        //
        }
        else if (!value && speechToText.IsListening)
        {
            speechToText.StopListening();
        }
    }
}

private void StartRecording()
{
    if (recordingRoutine == 0)
    {
        UnityObjectUtil.StartDestroyQueue();
        recordingRoutine = Runnable.Run(RecordingHandler());
    }
}

private void StopRecording()
{
    if (recordingRoutine != 0)
    {
        Microphone.End(microphoneID);
        Runnable.Stop(recordingRoutine);
        recordingRoutine = 0;
    }
}

private void OnSpeechError(string error)
{
    Active = false;

    Log.Debug("ExampleStreaming.OnError()", "Error! {0}", error);
}

private IEnumerator RecordingHandler()
{
    recording = Microphone.Start(microphoneID, true, recordingBufferSize, recordingHZ);
    yield return null;      // let _recordingRoutine get set..

    if (recording == null)
    {
        StopRecording();
        yield break;
    }

    bool bFirstBlock = true;
    int midPoint = recording.samples / 2;
    float[] samples = null;

    while (recordingRoutine != 0 && recording != null)
    {
        int writePos = Microphone.GetPosition(microphoneID);
        if (writePos > recording.samples || !Microphone.IsRecording(microphoneID))
        {
            Debug.Log("Microphone disconnected.");
            StopRecording();
            yield break;
        }

        if ((bFirstBlock && writePos >= midPoint) || (!bFirstBlock && writePos < midPoint))
        {
            // front block is recorded, make a RecordClip and pass it onto our callback.
            samples = new float[midPoint];
            recording.GetData(samples, bFirstBlock ? 0 : midPoint);

            AudioData record = new AudioData();
            record.MaxLevel = Mathf.Max(Mathf.Abs(Mathf.Min(samples)), Mathf.Max(samples));
            record.Clip = AudioClip.Create("Recording", midPoint, recording.channels, recordingHZ, false);
            record.Clip.SetData(samples, 0);

            speechToText.OnListen(record);

            bFirstBlock = !bFirstBlock;
        }
        else
        {
            // calculate the number of samples remaining until we ready for a block of audio, 
            // and wait that amount of time it will take to record.
            int remaining = bFirstBlock ? (midPoint - writePos) : (recording.samples - writePos);
            float timeRemaining = (float)remaining / (float)recordingHZ;

            yield return new WaitForSeconds(timeRemaining);
        }
    }

    yield break;
}

private void OnSpeechRecognize(SpeechRecognitionEvent result, Dictionary<string, object> customData)
{
    if (result != null && result.results.Length > 0)
    {
        foreach (var res in result.results)
        {
            foreach (var alt in res.alternatives)
            {

                string text = string.Format("{0} ({1}, {2:0.00})\n", alt.transcript, res.final ? "Final" : "Interim", alt.confidence);

                if (speechToTextField != null)
                {
                    speechToTextField.text = text;
                }

                if (res.final)
                {
                    if (characterState == SocialState.listening)
                    {
                        Debug.Log("WATSON | Speech to text recorded: \n" + alt.transcript);
                        StartCoroutine(Message(alt.transcript));
                    }
                }
                else
                {
                    if (characterState == SocialState.idle)
                    {
                        characterState = SocialState.listening;
                    }
                }
            }
        }
    }
}


// text to speech
private IEnumerator Synthesize(string text)
{
    Debug.Log("WATSON CALL | Synthesize input: \n" + text);

    textToSpeech.Voice = voiceType;
    bool doSynthesize = textToSpeech.ToSpeech(HandleSynthesizeCallback, OnFail, text, true);

    if (doSynthesize)
    {
        StartCoroutine(Analyze(text));
        saying = text;
        characterState = SocialState.talking;
    }
    yield return null;
}

void HandleSynthesizeCallback(AudioClip clip, Dictionary<string, object> customData = null)
{
    if (Application.isPlaying && clip != null)
    {
        voiceSource.clip = clip;
        voiceSource.Play();
    }
}

// conversation
private IEnumerator Message(string text)
{
    Debug.Log("WATSON | Conversation input: \n" + text);

    MessageRequest messageRequest = new MessageRequest()
    {
        input = new Dictionary<string, object>()
        {
            { "text", text }
        },
        context = conversationContext
    };
    bool doMessage = conversation.Message(HandleMessageCallback, OnFail, workspaceId, messageRequest);

    if (doMessage)
    {
        characterState = SocialState.thinking;

        if (conversationInputField != null)
        {
            conversationInputField.text = text;
        }
    }

    yield return null;
}

void HandleMessageCallback(object resp, Dictionary<string, object> customData)
{
    object _tempContext = null;
    (resp as Dictionary<string, object>).TryGetValue("context", out _tempContext);

    if (_tempContext != null)
        conversationContext = _tempContext as Dictionary<string, object>;
    string contextList = conversationContext.ToString();

    Dictionary<string, object> dict = Json.Deserialize(customData["json"].ToString()) as Dictionary<string, object>;
    Dictionary<string, object> output = dict["output"] as Dictionary<string, object>;
    Debug.Log("JSON INFO: " + customData["json"].ToString());

    // Send new/update context variables to the Watson Conversation Service
    if (weather.temperatureCity != null && !conversationContext.ContainsKey("temperature"))
    {
        string currentTemperature = weather.temperatureNumber.ToString();
        conversationContext.Add("temperature", currentTemperature);
    }
    else if (conversationContext.ContainsKey("temperature"))
    {
        string currentTemperature = weather.temperatureNumber.ToString();
        conversationContext.Remove("temperature");
        conversationContext.Add("temperature", currentTemperature);
        //Debug.Log("Current Temperature: " + currentTemperature);
    }

    // $ call context variables
    var context = dict["context"] as Dictionary<string, object>;
    if (context["destination_city"] != null)
    {
        destinationCity = context["destination_city"].ToString();
        Debug.Log("Destination city: " + destinationCity);
    }
    if (context["departure_city"] != null)
    {
        departureCity = context["departure_city"].ToString();
    }

    List<object> text = output["text"] as List<object>;
    string answer = text[0].ToString(); //Geeft alleen de eerste response terug

    Debug.Log("WATSON | Conversation output: \n" + answer);

    if (conversationOutputField != null)
    {
        conversationOutputField.text = answer;
    }

    fsData fsdata = null;
    fsResult r = _serializer.TrySerialize(resp.GetType(), resp, out fsdata);
    if (!r.Succeeded)
    {
        throw new WatsonException(r.FormattedMessages);
    }

    //convert fsdata to MessageResponse
    MessageResponse messageResponse = new MessageResponse();
    object obj = messageResponse;
    r = _serializer.TryDeserialize(fsdata, obj.GetType(), ref obj);
    if (!r.Succeeded)
    {
        throw new WatsonException(r.FormattedMessages);
    }

    if (resp != null)
    {
        //Recognize intents & entities
        if (messageResponse.intents.Length > 0 && messageResponse.entities.Length > 0)
        {
            string intent = messageResponse.intents[0].intent;
            string entity = messageResponse.entities[0].entity;
            string literalEntity = messageResponse.entities[0].value;
            if (entity == "city")
            {
                literalEntityCity = literalEntity;
            }
            if (intent == "weather" && entity == "city")
            {
                literalEntityCity = literalEntity;
            }
        }
        if (messageResponse.intents.Length > 0)
        {
            string intent = messageResponse.intents[0].intent;
            //Debug.Log("Intent: " + intent);                           //intent name
        }
        if (messageResponse.entities.Length > 0)
        {
            string entity = messageResponse.entities[0].entity;
            //Debug.Log("Entity: " + entity);                             //entity name
            string literalEntity = messageResponse.entities[0].value;
            //Debug.Log("Entity Literal: " + literalEntity);                //literal spoken entity
            if (entity == "city")
            {
                literalEntityCity = literalEntity;
            }
        }
    }

    StartCoroutine(Synthesize(answer));
}
}

Answer 1

您被问到的问题相当复杂。我相信如果你训练一个模型，它应该使用 Watson 的工具，而不是与 Unity 相关的东西。

但是，您在 Unity 中可以做的是更正返回词。也就是说，如果您希望只获得城市名称，则可以下载所有城市的列表，假设居民超过 100.000(您可以在 Internet 上找到它)，然后检查返回的单词是否在这个列表。例如:

http://download.geonames.org/export/dump/

如果不是，您可以认为它未被 Watson 检测到，因此您可以使用 Levenshtein 距离之类的东西来更正返回的单词。检查this

基本上，该算法试图找出两个词的不同之处。可以使用其他算法来检查与列表中最相似的给定单词。你可能会从here得到一些想法。或其他 one