sherpa_onnx提供语音复刻功能了

sherpa_onnx以前只能使用已有的语音模型，现在可以使用自己的声音了。

参考sherpa-onnx/pascal-api-examples/tts/zipvoice-zh-en.pas at master · k2-fsa/sherpa-onnx · GitHub

{ Copyright (c) 2026 Xiaomi Corporation }
program zipvoice_zh_en;
{
This file shows how to use the text to speech API of sherpa-onnx
with ZipVoice TTS models.

It generates speech from text and saves it to a wave file.

Please visit
https://k2-fsa.github.io/sherpa/onnx/tts/zipvoice.html
to download the model.
}

{$mode objfpc}

uses
ctypes,
SysUtils,
sherpa_onnx;

function ProgressCallback(Samples: pcfloat; N: cint32; P: cfloat;
Arg: Pointer): cint32; cdecl;
begin
WriteLn(Format('Progress: %.2f%%, samples: %d', [P * 100.0, N]));
Result := 1;
end;

function GetOfflineTts: TSherpaOnnxOfflineTts;
var
Config: TSherpaOnnxOfflineTtsConfig;
begin
Config := Default(TSherpaOnnxOfflineTtsConfig);
Config.Model.ZipVoice.Tokens := './sherpa-onnx-zipvoice-distill-int8-zh-en-emilia/tokens.txt';
Config.Model.ZipVoice.Encoder := './sherpa-onnx-zipvoice-distill-int8-zh-en-emilia/encoder.int8.onnx';
Config.Model.ZipVoice.Decoder := './sherpa-onnx-zipvoice-distill-int8-zh-en-emilia/decoder.int8.onnx';
Config.Model.ZipVoice.Vocoder := './vocos_24khz.onnx';
Config.Model.ZipVoice.DataDir := './sherpa-onnx-zipvoice-distill-int8-zh-en-emilia/espeak-ng-data';
Config.Model.ZipVoice.Lexicon := './sherpa-onnx-zipvoice-distill-int8-zh-en-emilia/lexicon.txt';
Config.Model.ZipVoice.FeatScale := 0.1;
Config.Model.ZipVoice.Tshift := 0.5;
Config.Model.ZipVoice.TargetRms := 0.1;
Config.Model.ZipVoice.GuidanceScale := 1.0;
Config.Model.NumThreads := 2;
Config.Model.Debug := True;
Config.MaxNumSentences := 1;

Result := TSherpaOnnxOfflineTts.Create(Config);
end;

var
Tts: TSherpaOnnxOfflineTts;
GenerationConfig: TSherpaOnnxGenerationConfig;
Wave: TSherpaOnnxWave;
WaveFilename: AnsiString;
Audio: TSherpaOnnxGeneratedAudio;
Text: AnsiString;
ReferenceText: AnsiString;

begin
Tts := GetOfflineTts;

WriteLn('There are ', Tts.GetNumSpeakers, ' speakers');

Text := '小米的价值观是真诚, 热爱. 真诚，就是不欺人也不自欺. 热爱, 就是全心投入并享受其中.';
ReferenceText := '那还是三十六年前, 一九八七年. 我呢考上了武汉大学的计算机系.';

WaveFilename := './sherpa-onnx-zipvoice-distill-int8-zh-en-emilia/test_wavs/leijun-1.wav';
Wave := SherpaOnnxReadWave(WaveFilename);
GenerationConfig := Default(TSherpaOnnxGenerationConfig);
GenerationConfig.SilenceScale := 0.2;
GenerationConfig.Speed := 1.0;
GenerationConfig.Sid := 0;
GenerationConfig.ReferenceAudio := Wave.Samples;
GenerationConfig.ReferenceAudioLen := Length(Wave.Samples);
GenerationConfig.ReferenceSampleRate := Wave.SampleRate;
GenerationConfig.ReferenceText := ReferenceText;
GenerationConfig.NumSteps := 4;
GenerationConfig.Extra := '{"min_char_in_sentence": "10"}';

Audio := Tts.Generate(Text, GenerationConfig, @ProgressCallback, NIL);
SherpaOnnxWriteWave('./zipvoice-zh-en.wav', Audio.Samples, Audio.SampleRate);
WriteLn('Saved to ./zipvoice-zh-en.wav');

FreeAndNil(Tts);
end.

还是复刻得比较像的。不过比前面提到的生成自己的onnx语音模型转换时要慢一点。原声7s，输出一个25s的音频我的I7700k大约18s，用gpu（1050）大约12s。