Fabio
c36510c21e
Some checks failed
Linux Build / Linux cpu (push) Has been cancelled
Linux Build / Linux cpu-clang (push) Has been cancelled
Linux Build / Linux vulkan (push) Has been cancelled
mac-build / macOS CPU (push) Has been cancelled
Nix Build / Nix (cpu) (push) Has been cancelled
Nix Build / Nix (vulkan) (push) Has been cancelled
windows-build / Windows CPU (push) Has been cancelled
27 lines
1.4 KiB
C++
27 lines
1.4 KiB
C++
#pragma once
|
|
#include "engine/framework/audio/conversion.h"
|
|
#include "engine/models/chatterbox/components.h"
|
|
#include <algorithm>
|
|
#include <stdexcept>
|
|
|
|
namespace engine::community_models::kitten_tts2 {
|
|
|
|
// Shared by live cloning and the offline preset preparer. The full decoder
|
|
// repeats short references to six seconds, then uses at most ten seconds.
|
|
inline models::chatterbox::EmbedReferenceOutputs prepare_decoder_reference(
|
|
runtime::AudioBuffer wav24,
|
|
const models::chatterbox::S3TokenizerComponent & tokenizer,
|
|
const models::chatterbox::CAMPPlusEncoderComponent & camp) {
|
|
if (wav24.sample_rate != 24000 || wav24.channels != 1 || wav24.samples.empty())
|
|
throw std::runtime_error("Kitten decoder reference must be non-empty mono 24 kHz audio");
|
|
const size_t original = wav24.samples.size();
|
|
const size_t repeats = (144000 + original - 1) / original;
|
|
const size_t length = std::min<size_t>(240000, original * std::max<size_t>(1, repeats));
|
|
wav24.samples.resize(length);
|
|
for (size_t i = original; i < length; ++i) wav24.samples[i] = wav24.samples[i % original];
|
|
runtime::AudioBuffer decoder16{16000, 1, audio::resample_mono_torchaudio_sinc_hann(
|
|
wav24.samples, 24000, 16000, audio::torchaudio_sinc_hann_float32_options())};
|
|
return tokenizer.embed_reference_from_wavs(camp, wav24, decoder16);
|
|
}
|
|
|
|
} // namespace engine::community_models::kitten_tts2
|