微软语音识别:结果与信心得分交替？

Question

我刚开始使用Microsoft.Speech识别器(使用11)，我试图让它从一个简单的语法中输出n个最佳识别匹配，以及每个语法的可信度分数。

根据文档(和前面提到的in the answer to this question)，人们应该能够使用e.Result.Alternates来访问得分最高的单词以外的其他可识别的单词。然而，即使在将置信度拒绝阈值重置为0之后(这意味着没有任何结果被拒绝)，我仍然只能得到一个结果，并且没有备选结果(尽管SpeechHypothesized事件表明至少有一个其他单词似乎在某个时候被识别为非零置信度)。

我的问题是:有人能向我解释为什么我只得到一个公认的词，即使置信度被设定为零？我怎样才能得到其他可能的比赛和他们的信心得分？我在这里错过了什么？

下面是我的密码。预先感谢任何能帮忙的人:)

在下面的示例中，识别器被发送一个单词"news“的wav文件，并且必须从类似的单词("noose"，"newts")中选择。我想为每个单词提取一个识别者的信心分数的列表(它们都应该是非零的)，即使结果它只会返回最好的一个(“新闻”)。

using System;
using System.Collections.Generic;
using System.Linq;
using System.Text;
using System.Threading.Tasks;
using Microsoft.Speech.Recognition;

namespace SimpleRecognizer
{
    class Program
    {
        static readonly string[] settings = new string[] {
            "CFGConfidenceRejectionThreshold",
            "HighConfidenceThreshold", 
            "NormalConfidenceThreshold",
            "LowConfidenceThreshold"};

        static void Main(string[] args)
        {
            // Create a new SpeechRecognitionEngine instance.
            SpeechRecognitionEngine sre = new SpeechRecognitionEngine(); //en-US SRE

            // Configure the input to the recognizer.
            sre.SetInputToWaveFile(@"C:\Users\Anjana\Documents\news.wav");

            // Display Recognizer Settings (Confidence Thresholds)
            ListSettings(sre);

            // Set Confidence Threshold to Zero (nothing should be rejected)
            sre.UpdateRecognizerSetting("CFGConfidenceRejectionThreshold", 0);
            sre.UpdateRecognizerSetting("HighConfidenceThreshold", 0);
            sre.UpdateRecognizerSetting("NormalConfidenceThreshold", 0);
            sre.UpdateRecognizerSetting("LowConfidenceThreshold", 0);

            // Display New Recognizer Settings
            ListSettings(sre);

            // Build a simple Grammar with three choices
            Choices topics = new Choices();
            topics.Add(new string[] { "news", "newts", "noose" });
            GrammarBuilder gb = new GrammarBuilder();
            gb.Append(topics);
            Grammar g = new Grammar(gb);
            g.Name = "g";

            // Load the Grammar
            sre.LoadGrammar(g);

            // Register handlers for Grammar's SpeechRecognized Events
            g.SpeechRecognized += new EventHandler<SpeechRecognizedEventArgs>(gram_SpeechRecognized);

            // Register a handler for the recognizer's SpeechRecognized event.
            sre.SpeechRecognized += new EventHandler<SpeechRecognizedEventArgs>(sre_SpeechRecognized);

            // Register Handler for SpeechHypothesized
            sre.SpeechHypothesized += new EventHandler<SpeechHypothesizedEventArgs>(sre_SpeechHypothesized);

            // Start recognition.
            sre.Recognize();

            Console.ReadKey(); //wait to close

        }
        static void gram_SpeechRecognized(object sender, SpeechRecognizedEventArgs e)
        {
            Console.WriteLine("\nNumber of Alternates from Grammar {1}: {0}", e.Result.Alternates.Count.ToString(), e.Result.Grammar.Name);
            foreach (RecognizedPhrase phrase in e.Result.Alternates)
            {
                Console.WriteLine(phrase.Text + ", " + phrase.Confidence);
            }
        }
        static void sre_SpeechRecognized(object sender, SpeechRecognizedEventArgs e)
        {
            Console.WriteLine("\nSpeech recognized: " + e.Result.Text + ", " + e.Result.Confidence);
            Console.WriteLine("Number of Alternates from Recognizer: {0}", e.Result.Alternates.Count.ToString());
            foreach (RecognizedPhrase phrase in e.Result.Alternates)
            {
                Console.WriteLine(phrase.Text + ", " + phrase.Confidence);
            }
        }
        static void sre_SpeechHypothesized(object sender, SpeechHypothesizedEventArgs e)
        {
            Console.WriteLine("Speech from grammar {0} hypothesized: {1}, {2}", e.Result.Grammar.Name, e.Result.Text, e.Result.Confidence);
        }
        private static void ListSettings(SpeechRecognitionEngine recognizer)
        {
            foreach (string setting in settings)
            {
                try
                {
                    object value = recognizer.QueryRecognizerSetting(setting);
                    Console.WriteLine("  {0,-30} = {1}", setting, value);
                }
                catch
                {
                    Console.WriteLine("  {0,-30} is not supported by this recognizer.",
                      setting);
                }
            }
            Console.WriteLine();
        }
    }
}

这提供了以下输出：

Original recognizer settings:
  CFGConfidenceRejectionThreshold = 20
  HighConfidenceThreshold        = 80
  NormalConfidenceThreshold      = 50
  LowConfidenceThreshold         = 20

Updated recognizer settings:
  CFGConfidenceRejectionThreshold = 0
  HighConfidenceThreshold        = 0
  NormalConfidenceThreshold      = 0
  LowConfidenceThreshold         = 0

Speech from grammar g hypothesized: noose, 0.2214646
Speech from grammar g hypothesized: news, 0.640804

Number of Alternates from Grammar g: 1
news, 0.9208503

Speech recognized: news, 0.9208503
Number of Alternates from Recognizer: 1
news, 0.9208503

我还试着用一个单独的短语来实现这一点(而不是一个有三个选择的短语)，甚至每个单词/短语都有一个单独的语法。结果基本上是一样的:只有一个“交替”。

Answer 1

我相信这是SAPI允许您请求SR引擎不支持的东西的另一个地方。

Microsoft.Speech.Recognition和System.Speech.Recognition都使用底层的SAPI接口来完成他们的工作；唯一的区别是使用了哪个SR引擎。(Microsoft.Speech.Recognition使用服务器引擎；System.Speech.Recognition使用桌面引擎。)

交替词主要是为听写而设计的，而不是上下文无关的语法。您总是可以为CFG获得一个备用的，但是替代的生成代码看起来不会扩展CFG的备用代码。

不幸的是，Microsoft.Speech.Recognition引擎不支持听写。(然而，它的工作音频质量要低得多，而且不需要培训。)