接入 SenseVoice模型

VideoAnalysis/Program.cs

@@ -49,6 +49,7 @@ namespace Learn.VideoAnalysis
             //初始化 插件
             Speaker.Init();
             RedisExpand.Init();
+            SenseVoice.Init();
 
 
             builder.Services.AddScoped(sp =>

VideoAnalysis/appsettings.json

@@ -19,7 +19,8 @@
     "ChatGpt": {
       "KIMI": {
         "Host": "https://api.moonshot.cn",
-        "ApiKey": "sk-CNYJdRHgJsgtgw1Q8GhQ5ayXuFPVLSk5bduOF4l2FMvI5lUo"
+        //"ApiKey": "sk-CNYJdRHgJsgtgw1Q8GhQ5ayXuFPVLSk5bduOF4l2FMvI5lUo"
+        "ApiKey": "sk-8BvvhESZIkgUbiaaJhglPxFa4o2X9H3xEv9lXELrWWwGxHWY"
       }
     },
     "DB": {

VideoAnalysisCore/AICore/ChatGPT/KIMI/KIMI_GPT.cs

@@ -67,7 +67,7 @@ namespace VideoAnalysisCore.AICore.ChatGPT.KIMI
                 criteriaBuilder.Append("|");
             }
 
-            var resFormat = "问题编号:int,结果:array|bool,问题解释:string";
+            var resFormat = "[{问题编号:int,结果:array|bool,问题解释:string}]";
             var postMessages = 
                 $"以下是一段音频的字幕,分析这段字幕(格式 说话人:开始秒:结束秒:内容|下一段字幕)." +
                 $"来简明的回答提出的问题 问题列表 {criteriaBuilder} " +
@@ -80,15 +80,14 @@ namespace VideoAnalysisCore.AICore.ChatGPT.KIMI
             var modelId = reqTokenCount > 32 * 1000 ? "moonshot-v1-128k" : "moonshot-v1-32k";
             var chatRep = new ChatReq
             {
-                max_tokens =1000 * 31,
+                max_tokens = reqTokenCount * 2,
                 temperature = 0.3,
                 frequency_penalty = 0,
                 presence_penalty = 0,
                 model = modelId,
                 messages = new List<MessagesItem>(){
                     new MessagesItem(postMessages,"system"),
-                    //todo 规定返回json格式
+                    new MessagesItem(postMessages,"assistant"),
-                    //new MessagesItem(postMessages,"assistant"),
                 }
             };
             var chatResp = await moonshotClient.Chat(chatRep);

VideoAnalysisCore/AICore/SherpaOnnx/SenseVoice.cs

@@ -0,0 +1,148 @@
+using Microsoft.Extensions.Options;
+using SherpaOnnx;
+using System;
+using System.Collections.Generic;
+using System.Linq;
+using System.Text;
+using System.Threading.Tasks;
+using VideoAnalysisCore.AICore.Whisper;
+using VideoAnalysisCore.Common;
+
+namespace VideoAnalysisCore.AICore.SherpaOnnx
+{
+    public  class SenseVoice
+    {
+        static OfflineRecognizer OR =default!;
+        static VoiceActivityDetector VAD =default!;
+        static VadModelConfig VADModelConfig = default!;
+        /// <summary>
+        /// 初始化 SenseVoice
+        /// </summary>
+        /// <param name="speakerNumber"></param>
+        /// <param name="threshold"></param>
+        public static void Init(int speakerNumber = 0, double threshold = 0.6)
+        {
+            Console.WriteLine("初始化 SenseVoice");
+            OfflineRecognizerConfig config = new OfflineRecognizerConfig();
+            //采样率
+            config.FeatConfig.SampleRate = 16000;
+            //用于训练模型的特征维度
+            config.FeatConfig.FeatureDim = 80;
+            //Path to tokens.txt
+            config.ModelConfig.Tokens = Path.Combine(AppCommon.AIModelFile, "sherpa-onnx-sense-voice-24-07-17", "tokens.txt");
+            //SenseVoice 模型
+            config.ModelConfig.SenseVoice.Model = Path.Combine(AppCommon.AIModelFile, "sherpa-onnx-sense-voice-24-07-17", "model.onnx");
+            //1 使用逆文本规范化处理感官语音。
+            config.ModelConfig.SenseVoice.UseInverseTextNormalization =1;
+            //模型类型
+            config.ModelConfig.ModelType = string.Empty;
+
+            #region 有效的解码方法
+            //贪婪搜索[greedy_search] 改进的波束搜索 [modified_beam_search]
+            //贪婪搜索
+            config.DecodingMethod = "greedy_search";
+
+            ////改进的波束搜索
+            //config.DecodingMethod = "modified_beam_search";
+            ////仅在 --decoding--method 为 [波束搜索]modified_beam_search 时使用。
+            ////它指定搜索过程中要保留的活动路径数
+            //config.MaxActivePaths =4;
+            #endregion
+
+            //热词目录
+            config.HotwordsFile = string.Empty;
+            //热词得分
+            config.HotwordsScore =1.5f ;
+            //反转文本规范化规则 fst 的路径
+            config.RuleFsts =  string.Empty;
+
+            config.ModelConfig.Debug = 0;
+
+            OR = new OfflineRecognizer(config);
+
+            VADModelConfig = new VadModelConfig();
+            VADModelConfig.SileroVad.Model = Path.Combine(AppCommon.AIModelFile, "sherpa-onnx-sense-voice-24-07-17", "silero_VAD.onnx");
+            VADModelConfig.Debug = 0;
+            //缓冲区大小
+            VAD = new VoiceActivityDetector(VADModelConfig, 60);
+        }
+        /// <summary>
+        /// 获取语音字幕
+        /// </summary>
+        /// <param name="task"></param>
+        /// <returns></returns>
+        public static async Task RunTask(string task)
+        {
+            var filePath = Path.Combine(task.LocalPath(), task + ".wav");
+            if (string.IsNullOrEmpty(filePath) || !File.Exists(filePath))
+                throw new Exception("task 音频路径未找到");
+
+            string testWaveFilename = filePath;
+            WaveReader reader = new WaveReader(testWaveFilename);
+
+            int numSamples = reader.Samples.Length;
+            int windowSize = VADModelConfig.SileroVad.WindowSize;
+            int sampleRate = VADModelConfig.SampleRate;
+            int numIter = numSamples / windowSize;
+
+            var res = new List<SenseVoiceRes>(500);
+            for (int i = 0; i != numIter; ++i)
+            {
+                int start = i * windowSize;
+                float[] samples = new float[windowSize];
+                Array.Copy(reader.Samples, start, samples, 0, windowSize);
+                VAD.AcceptWaveform(samples);
+                //是否检测到语音
+                if (VAD.IsSpeechDetected())
+                {
+                    while (!VAD.IsEmpty())
+                    {
+                        //获取最新的发言片段
+                        SpeechSegment segment = VAD.Front();
+                        float startTime = segment.Start / (float)sampleRate;
+                        float duration = segment.Samples.Length / (float)sampleRate;
+                        OfflineStream stream = OR.CreateStream();
+                        stream.AcceptWaveform(sampleRate, segment.Samples);
+                        OR.Decode(stream);
+                        if (!string.IsNullOrEmpty(stream.Result.Text))
+                        {
+                            res.Add(new() 
+                            { 
+                                Text = stream.Result.Text,
+                                Start= startTime,
+                                End = startTime + duration });
+                        }
+                        VAD.Pop();
+                    }
+                }
+            }
+            VAD.Flush();
+
+            while (!VAD.IsEmpty())
+            {
+                SpeechSegment segment = VAD.Front();
+                float startTime = segment.Start / (float)sampleRate;
+                float duration = segment.Samples.Length / (float)sampleRate;
+
+                OfflineStream stream = OR.CreateStream();
+                stream.AcceptWaveform(sampleRate, segment.Samples);
+                OR.Decode(stream);
+                if (!string.IsNullOrEmpty(stream.Result.Text))
+                {
+                    res.Add(new()
+                    {
+                        Text = stream.Result.Text,
+                        Start = startTime,
+                        End = startTime + duration
+                    });
+                }
+
+                VAD.Pop();
+            }
+
+            await RedisExpand.Redis.HMSetAsync(RedisExpandKey.Task(task), "Captions", res);
+            RedisExpand.InsertChannel(Enum.RedisChannelEnum.ParsingSpeaker, task);
+
+        }
+    }
+}

VideoAnalysisCore/AICore/SherpaOnnx/SherpaOnnxDto.cs

@@ -0,0 +1,29 @@
+using Whisper.net;
+
+namespace VideoAnalysisCore.AICore.Whisper
+{
+    /// <summary>
+    ///  字幕识别 结果
+    /// </summary>
+    public class SenseVoiceRes
+    {
+        public SenseVoiceRes()
+        {
+                
+        }
+        /// <summary>
+        /// 文本
+        /// </summary>
+        public string Text { get; set; } = string.Empty;
+        /// <summary>
+        /// 开始时间
+        /// </summary>
+
+        public float Start { get; set; }
+        /// <summary>
+        /// 结束时间
+        /// </summary>
+
+        public float End { get; set; }
+    }
+}

VideoAnalysisCore/AICore/SherpaOnnx/Speaker.cs

@@ -22,6 +22,7 @@ namespace VideoAnalysisCore.AICore.SherpaOnnx
         /// <param name="threshold"></param>
         public static void Init(int speakerNumber = 0, double threshold = 0.6)
         {
+            Console.WriteLine("初始化 Speaker");
             var config = new OfflineSpeakerDiarizationConfig();
             //Pyannote模型地址
             config.Segmentation.Pyannote.Model = Path.Combine(AppCommon.AIModelFile, "sherpa-onnx-pyannote-segmentation-3-0", "model.onnx");
@@ -72,15 +73,15 @@ namespace VideoAnalysisCore.AICore.SherpaOnnx
         /// 总持续时间
         /// </summary>
         [JsonIgnore]
-        public decimal Total => End - Start;
+        public float Total => End - Start;
         /// <summary>
         /// 开始时间
         /// </summary>
-        public decimal Start { get; set; }
+        public float Start { get; set; }
         /// <summary>
         /// 结束时间
         /// </summary>
-        public decimal End { get; set; }
+        public float End { get; set; }
         /// <summary>
         /// 讲话人索引
         /// </summary>
@@ -99,8 +100,8 @@ namespace VideoAnalysisCore.AICore.SherpaOnnx
         /// <param name="sds"></param>
         public OfflineSpeakerRes(OfflineSpeakerDiarizationSegment sds)
         {
-            Start = (decimal)sds.Start;
+            Start = sds.Start;
-            End = (decimal)sds.End;
+            End =sds.End;
             SpeakerIndex = sds.Speaker;
         }
     }

VideoAnalysisCore/AICore/Whisper/WhisperDto.cs

@@ -7,6 +7,10 @@ namespace VideoAnalysisCore.AICore.Whisper
     /// </summary>
     public class WhisperResDto
     {
+        public WhisperResDto()
+        {
+                
+        }
         /// <summary>
         /// 
         /// </summary>
@@ -20,16 +24,16 @@ namespace VideoAnalysisCore.AICore.Whisper
         /// <summary>
         /// 文本
         /// </summary>
-        public string Text { get; } = string.Empty;
+        public string Text { get; set; } = string.Empty;
         /// <summary>
         /// 开始时间
         /// </summary>
 
-        public TimeSpan Start { get; }
+        public TimeSpan Start { get; set; }
         /// <summary>
         /// 结束时间
         /// </summary>
 
-        public TimeSpan End { get; }
+        public TimeSpan End { get; set; }
     }
 }

VideoAnalysisCore/Common/RedisExpand.cs

@@ -84,6 +84,7 @@ namespace VideoAnalysisCore.Common
         /// </summary>
         public static void Init()
         {
+            Console.WriteLine("初始化 redis");
             Redis.Serialize = obj => System.Text.Json.JsonSerializer.Serialize(obj);
             Redis.Deserialize = (json, type) => System.Text.Json.JsonSerializer.Deserialize(json, type);
             InitChannel();
@@ -125,7 +126,7 @@ namespace VideoAnalysisCore.Common
                 (msg) => { TouchChannel(RedisChannelEnum.SeparateAudio, msg, FFMPGEHandle.Audio2WAV16KAsync); });
 
             Redis.SubscribeList(RedisExpandKey.EnumKey(RedisChannelEnum.ParsingCaptions),
-                (msg) => { TouchChannel(RedisChannelEnum.ParsingCaptions, msg, WhisperHandle.RunTask); });
+                (msg) => { TouchChannel(RedisChannelEnum.ParsingCaptions, msg, SenseVoice.RunTask); });
             Redis.SubscribeList(RedisExpandKey.EnumKey(RedisChannelEnum.ParsingSpeaker),
                 (msg) => { TouchChannel(RedisChannelEnum.ParsingSpeaker, msg, Speaker.Run); });
             Redis.SubscribeList(RedisExpandKey.EnumKey(RedisChannelEnum.ChatModelAnalysis),

VideoAnalysisCore/VideoAnalysisCore.csproj

@@ -49,7 +49,7 @@
 		<PackageReference Include="Microsoft.Extensions.DependencyModel" Version="7.0.0" />
 		<PackageReference Include="Microsoft.Extensions.Http" Version="8.0.0" />
 		<PackageReference Include="Newtonsoft.Json" Version="13.0.3" />
-		<PackageReference Include="org.k2fsa.sherpa.onnx" Version="1.10.28" />
+		<PackageReference Include="org.k2fsa.sherpa.onnx" Version="1.10.30" />
 		<PackageReference Include="SqlSugar.IOC" Version="2.0.0" />
 		<PackageReference Include="SqlSugarCore" Version="5.1.4.170" />
 		<PackageReference Include="Whisper.net" Version="1.5.0" />