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
100 lines
3.9 KiB
C++
100 lines
3.9 KiB
C++
/*
|
|
* Rules behind the MOSS-TTS-Nano reference-voice cache.
|
|
*
|
|
* The cache exists because encoding a reference runs the audio tokenizer encoder, and
|
|
* a session that keeps one voice should pay that once. These checks are model-free on
|
|
* purpose: they run anywhere, including build configurations with no model downloaded.
|
|
*/
|
|
|
|
#include "engine/models/moss/moss_tts_nano/session.h"
|
|
|
|
#include <cstdio>
|
|
#include <optional>
|
|
#include <vector>
|
|
|
|
namespace {
|
|
|
|
namespace nano = engine::models::moss_tts_nano;
|
|
namespace runtime = engine::runtime;
|
|
|
|
int g_failures = 0;
|
|
|
|
void check(bool condition, const char * what) {
|
|
std::printf(" [%s] %s\n", condition ? "ok" : "FAIL", what);
|
|
if (!condition) {
|
|
++g_failures;
|
|
}
|
|
}
|
|
|
|
runtime::AudioBuffer make_audio(float bias) {
|
|
runtime::AudioBuffer buffer;
|
|
buffer.sample_rate = 48000;
|
|
buffer.channels = 2;
|
|
buffer.samples.resize(64);
|
|
for (size_t i = 0; i < buffer.samples.size(); ++i) {
|
|
buffer.samples[i] = bias + static_cast<float>(i) * 1e-4F;
|
|
}
|
|
return buffer;
|
|
}
|
|
|
|
nano::MossTTSNanoAudioCodes make_codes(int64_t codebooks) {
|
|
nano::MossTTSNanoAudioCodes codes;
|
|
codes.frames = 4;
|
|
codes.codebooks = codebooks;
|
|
codes.token_ids.assign(static_cast<size_t>(codes.frames * codebooks), 0);
|
|
return codes;
|
|
}
|
|
|
|
} // namespace
|
|
|
|
int main() {
|
|
const runtime::AudioBuffer voice_a = make_audio(0.10F);
|
|
const runtime::AudioBuffer voice_b = make_audio(0.20F);
|
|
|
|
std::optional<runtime::AudioBuffer> cached_audio;
|
|
std::optional<nano::MossTTSNanoAudioCodes> cached_codes;
|
|
|
|
std::printf("first prepare records the audio, encode happens on first use\n");
|
|
nano::invalidate_reference_codes_if_changed(cached_audio, cached_codes, voice_a);
|
|
check(cached_audio.has_value(), "audio recorded");
|
|
check(!cached_codes.has_value(), "prepare does not encode");
|
|
check(!nano::reference_codes_reusable(cached_audio, cached_codes, voice_a, 16),
|
|
"without codes nothing is reusable");
|
|
|
|
std::printf("after the first encode the same voice is reused\n");
|
|
cached_codes = make_codes(16);
|
|
check(nano::reference_codes_reusable(cached_audio, cached_codes, voice_a, 16),
|
|
"same audio and codebook count -> reusable");
|
|
check(!nano::reference_codes_reusable(cached_audio, cached_codes, voice_a, 8),
|
|
"different codebook count -> not reusable");
|
|
check(!nano::reference_codes_reusable(cached_audio, cached_codes, voice_b, 16),
|
|
"different audio -> not reusable");
|
|
|
|
std::printf("preparing the same voice again keeps the codes\n");
|
|
nano::invalidate_reference_codes_if_changed(cached_audio, cached_codes, voice_a);
|
|
check(cached_codes.has_value(), "codes kept");
|
|
check(nano::reference_codes_reusable(cached_audio, cached_codes, voice_a, 16), "still reusable");
|
|
|
|
std::printf("changing the voice invalidates them\n");
|
|
nano::invalidate_reference_codes_if_changed(cached_audio, cached_codes, voice_b);
|
|
check(!cached_codes.has_value(), "codes dropped, so the new voice cannot reuse the old codes");
|
|
cached_codes = make_codes(16);
|
|
check(nano::reference_codes_reusable(cached_audio, cached_codes, voice_b, 16), "the new voice is cached");
|
|
|
|
std::printf("near misses are not treated as the same voice\n");
|
|
runtime::AudioBuffer tweaked = voice_b;
|
|
tweaked.samples.back() += 0.5F;
|
|
check(!nano::reference_codes_reusable(cached_audio, cached_codes, tweaked, 16),
|
|
"one changed sample -> not reusable");
|
|
runtime::AudioBuffer mono = voice_b;
|
|
mono.channels = 1;
|
|
check(!nano::reference_codes_reusable(cached_audio, cached_codes, mono, 16),
|
|
"different channel count -> not reusable");
|
|
runtime::AudioBuffer resampled = voice_b;
|
|
resampled.sample_rate = 44100;
|
|
check(!nano::reference_codes_reusable(cached_audio, cached_codes, resampled, 16),
|
|
"different sample rate -> not reusable");
|
|
|
|
std::printf("%s\n", g_failures == 0 ? "PASS: moss_tts_nano reference cache" : "FAILED");
|
|
return g_failures == 0 ? 0 : 1;
|
|
}
|