Learn.VideoAnalysis/VideoAnalysisCore/AICore/SherpaOnnx/SenseVoice.cs

using Dm.util;
using Microsoft.Extensions.DependencyInjection;
using Microsoft.Extensions.Options;
using SherpaOnnx;
using SqlSugar.IOC;
using System;
using System.Collections.Generic;
using System.Diagnostics;
using System.IO;
using System.Linq;
using System.Text;
using System.Text.Json;
using System.Text.RegularExpressions;
using System.Threading.Tasks;
using VideoAnalysisCore.Common;
using VideoAnalysisCore.Model;
using VideoAnalysisCore.Model.Enum;

namespace VideoAnalysisCore.AICore.SherpaOnnx
{
    public static class SenseVoiceExpand 
    {

        /// <summary>
        /// 添加 SenseVoice 语音转文字
        /// </summary>
        /// <param name="services"></param>
        public static void AddSenseVoiceExpand(this IServiceCollection services)
        {
            services.AddSingleton<SenseVoice>();
        }
    }
    public class SenseVoice
    {
        public static OfflineRecognizer OR = default!; 
        private readonly IServiceProvider serviceProvider;

        public static OfflineRecognizer OR1 = default!;
        //测试用
        public static List<(string z1,string z2)> cachedValue = new List<(string z1, string z2)>();


        public SenseVoice(RedisManager redisManager, IServiceProvider serviceProvider)
        {
            this.serviceProvider = serviceProvider;
        }

        /// <summary>
        /// 初始化 SenseVoice 
        /// </summary>
        /// <param name="numThreads">默认6线程</param>
        /// <param name="useGPU">是否使用gpu 报错请看安装CUDA环境<see cref="https://k2-fsa.github.io/sherpa/onnx/pretrained_models/whisper/large-v3.html#run-with-gpu-float32"/></param>
        public void Init(int numThreads = 6, bool useGPU = false, bool useHotwords = false)
        {
            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
            var AIModelVersion_270717 = "sherpa-onnx-sense-voice-24-07-17";
            config.ModelConfig.Tokens = Path.Combine(AppCommon.AIModelFile, AIModelVersion_270717, "tokens.txt");
            //SenseVoice 模型
            config.ModelConfig.SenseVoice.Model = Path.Combine(AppCommon.AIModelFile, AIModelVersion_270717, "model.onnx");
            //1 使用逆文本规范化处理感官语音 [控制标点符号生成]。
            config.ModelConfig.SenseVoice.UseInverseTextNormalization = 1;
            //反转文本规范化规则 fst 的路径
            //config.RuleFsts = Path.Combine(AppCommon.AIModelFile, "itn_subject_sx.fst");

            config.ModelConfig.SenseVoice.Language = "zh";
            //模型类型
            config.ModelConfig.ModelType = string.Empty;
            config.ModelConfig.NumThreads = numThreads;
            config.ModelConfig.Provider = "cpu";
            //需要使用GPU
            if (!useGPU)
                config.ModelConfig.Provider = "cuda";

            #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


#if DEBUG
            config.ModelConfig.Debug = 1;
#endif
            OR = new OfflineRecognizer(config);
            
            //var AIModelVersion_251217 = "sherpa-onnx-sense-voice-funasr-nano-2025-12-17";
            //OfflineRecognizerConfig config1 = new OfflineRecognizerConfig();
            //config1.FeatConfig.SampleRate = 16000;
            //config1.FeatConfig.FeatureDim = 80;
            //config1.ModelConfig.Tokens = Path.Combine(AppCommon.AIModelFile, AIModelVersion_251217, "tokens.txt");
            //config1.ModelConfig.SenseVoice.Model = Path.Combine(AppCommon.AIModelFile, AIModelVersion_251217, "model.onnx");
            ////1 使用逆文本规范化处理感官语音 [控制标点符号生成]。
            //config1.ModelConfig.SenseVoice.UseInverseTextNormalization = 1;
            //config1.ModelConfig.SenseVoice.Language = "zh";
            //config1.ModelConfig.ModelType = string.Empty;
            //config1.ModelConfig.NumThreads = numThreads;
            //config1.ModelConfig.Provider = "cpu";
            //config1.DecodingMethod = "greedy_search";
            //config1.ModelConfig.Debug = 1;
            //OR1 = new OfflineRecognizer(config: config1);
            //OR1 = FunASRNano.OR;

        }

        /// <summary>
        /// 获取语音字幕
        /// </summary>
        /// <param name="s"></param>
        /// <returns></returns>
        public List<SenseVoiceRes> RunTask(Stream s)
        {
            if (s is null) throw new Exception("音频路径 is null");
            if (OR is null) Init();
            return serviceProvider.GetRequiredService<SherpaVad>()
                .TaskHandle(new WaveReader(s), null, SoundHandle, SherpaVadVersion.silero_vad_v5);
        }
        /// <summary>
        /// 获取语音字幕
        /// </summary>
        /// <param name="task"></param>
        /// <returns></returns>
        public 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");
            if (OR is null) Init();
            serviceProvider.GetRequiredService<SherpaVad>()
               .TaskHandle(new WaveReader(filePath), task, SoundHandle, SherpaVadVersion.ten_vad_324);

            return Task.CompletedTask;
        }
  

        /// <summary>
        /// 获取语音字幕
        /// </summary>
        /// <param name="sampleRate">采样率</param>
        /// <param name="samples">采样值(样品)</param>
        /// <returns>结果流</returns>
        public OfflineStream SoundHandle(int sampleRate, float[] samples)
        {
            var stream = OR.CreateStream();
            stream.AcceptWaveform(sampleRate, samples);
            OR.Decode(stream);
            return stream;
        }

    }
}