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
154 lines
8.4 KiB
C++
154 lines
8.4 KiB
C++
// A voice preset's options reach the model as request options: they fill what the
|
|
// request leaves unset, for named and default presets alike, and the server never
|
|
// interprets them.
|
|
#include "server_protocol.h"
|
|
|
|
#include "engine/framework/io/json.h"
|
|
|
|
#include <filesystem>
|
|
#include <iostream>
|
|
#include <optional>
|
|
#include <stdexcept>
|
|
#include <string>
|
|
#include <unordered_map>
|
|
|
|
namespace {
|
|
|
|
namespace fs = std::filesystem;
|
|
using minitts::server::RuntimeVoicePreset;
|
|
using minitts::server::SpeechRequestContext;
|
|
|
|
void expect(bool condition, const std::string & message) {
|
|
if (!condition) {
|
|
throw std::runtime_error(message);
|
|
}
|
|
}
|
|
|
|
engine::runtime::TaskRequest speech(
|
|
const std::unordered_map<std::string, RuntimeVoicePreset> & presets,
|
|
const std::optional<RuntimeVoicePreset> & default_preset,
|
|
const std::unordered_map<std::string, std::string> & default_options,
|
|
const std::string & body_text) {
|
|
const fs::path base = fs::current_path();
|
|
const std::optional<fs::path> voice_dir;
|
|
return minitts::server::build_speech_request(
|
|
engine::io::json::parse(body_text),
|
|
SpeechRequestContext{base, voice_dir, presets, default_preset, true, true, true, default_options});
|
|
}
|
|
|
|
} // namespace
|
|
|
|
int main() {
|
|
try {
|
|
RuntimeVoicePreset narrator;
|
|
narrator.options = {
|
|
{"speaker_embedding_file", "/voices/narrator.speaker.safetensors"},
|
|
{"num_steps", "8"},
|
|
};
|
|
RuntimeVoicePreset with_text;
|
|
with_text.reference_text = "preset transcript";
|
|
with_text.options = {{"reference_text", "option transcript"}};
|
|
const std::unordered_map<std::string, RuntimeVoicePreset> presets{
|
|
{"narrator", narrator},
|
|
{"with_text", with_text},
|
|
};
|
|
const std::optional<RuntimeVoicePreset> no_default;
|
|
const std::unordered_map<std::string, std::string> no_defaults;
|
|
|
|
auto request = speech(presets, no_default, no_defaults, R"({"input":"x","voice":"narrator"})");
|
|
expect(request.options.at("speaker_embedding_file") == "/voices/narrator.speaker.safetensors",
|
|
"preset option did not reach the request");
|
|
expect(request.options.at("num_steps") == "8", "second preset option did not reach the request");
|
|
expect(!request.voice.has_value(), "an options-only preset should not add a voice");
|
|
|
|
// The request's own options win over the preset's.
|
|
request = speech(presets, no_default, no_defaults,
|
|
R"({"input":"x","voice":"narrator","options":{"num_steps":4}})");
|
|
expect(request.options.at("num_steps") == "4", "request option did not override the preset");
|
|
expect(request.options.at("speaker_embedding_file") == "/voices/narrator.speaker.safetensors",
|
|
"unrelated preset option was dropped");
|
|
|
|
// The preset wins over the model's default_request_options.
|
|
const std::unordered_map<std::string, std::string> defaults{{"num_steps", "32"}, {"cfg_scale", "2"}};
|
|
request = speech(presets, no_default, defaults, R"({"input":"x","voice":"narrator"})");
|
|
expect(request.options.at("num_steps") == "8", "default_request_options overrode the preset");
|
|
expect(request.options.at("cfg_scale") == "2", "default_request_options were dropped");
|
|
|
|
// The dedicated reference_text field keeps its meaning.
|
|
request = speech(presets, no_default, no_defaults, R"({"input":"x","voice":"with_text"})");
|
|
expect(request.options.at("reference_text") == "preset transcript",
|
|
"preset reference_text should win over the same key in its options");
|
|
|
|
// A default preset carries its options too, until the request names a voice.
|
|
const std::optional<RuntimeVoicePreset> default_preset = narrator;
|
|
request = speech(presets, default_preset, no_defaults, R"({"input":"x"})");
|
|
expect(request.options.at("speaker_embedding_file") == "/voices/narrator.speaker.safetensors",
|
|
"default preset options did not reach the request");
|
|
request = speech(presets, default_preset, no_defaults, R"({"input":"x","voice":"alba"})");
|
|
expect(request.options.count("speaker_embedding_file") == 0,
|
|
"default preset options leaked into a request that named another voice");
|
|
expect(request.voice.has_value() && request.voice->speaker.has_value() &&
|
|
request.voice->speaker->cached_voice_id == "alba",
|
|
"a voice that is not a preset should stay a model-native voice id");
|
|
|
|
RuntimeVoicePreset clone;
|
|
clone.audio = engine::runtime::AudioBuffer{24000, 1, {0.25F, -0.25F}};
|
|
clone.reference_text = "reference transcript";
|
|
for (const std::string body : {R"({"input":"x"})",
|
|
R"({"input":"x","voice":""})",
|
|
R"({"input":"x","voice":null})"}) {
|
|
request = speech(presets, clone, no_defaults, body);
|
|
expect(request.voice.has_value() && request.voice->speaker.has_value() &&
|
|
request.voice->speaker->audio.has_value(),
|
|
"omitted, empty, or null voice should retain the default reference");
|
|
expect(request.voice->speaker->audio->samples == clone.audio->samples &&
|
|
request.options.at("reference_text") == *clone.reference_text,
|
|
"default reference audio or transcript changed");
|
|
request = speech(presets, no_default, no_defaults, body);
|
|
expect(!request.voice.has_value(), "empty or null voice should not create a voice");
|
|
}
|
|
request = speech(presets, clone, no_defaults, R"({"input":"x","voice":"alba"})");
|
|
expect(request.voice->speaker->cached_voice_id == "alba" &&
|
|
!request.voice->speaker->audio.has_value(),
|
|
"nonempty voice must still override the default reference");
|
|
bool rejected = false;
|
|
try { (void)speech(presets, clone, no_defaults, R"({"input":"x","voice":123})"); }
|
|
catch (const std::runtime_error &) { rejected = true; }
|
|
expect(rejected, "non-string, non-null voice must still be rejected");
|
|
|
|
// Instruction aliases follow request > preset > model precedence.
|
|
for (const std::string key : {"instruct", "instruction"}) {
|
|
const std::string alias = key == "instruct" ? "instruction" : "instruct";
|
|
RuntimeVoicePreset styled = narrator;
|
|
styled.options[key] = "calm";
|
|
const std::unordered_map<std::string, RuntimeVoicePreset> styled_presets{{"styled", styled}};
|
|
const std::unordered_map<std::string, std::string> styled_defaults{{alias, "serious"}};
|
|
auto check = [&](const std::string & body, const std::optional<RuntimeVoicePreset> & selected,
|
|
const std::string & expected_key, const std::string & expected_value) {
|
|
const auto result = speech(styled_presets, selected, styled_defaults, body);
|
|
expect(result.options.at(expected_key) == expected_value, "instruction precedence changed");
|
|
expect(result.options.count(expected_key == "instruct" ? "instruction" : "instruct") == 0,
|
|
"lower-priority instruction alias survived");
|
|
expect(result.options.at("num_steps") == "8", "unrelated preset option changed");
|
|
expect(result.options.at("speaker_embedding_file") == "/voices/narrator.speaker.safetensors",
|
|
"unrelated voice option changed");
|
|
};
|
|
check(R"({"input":"x","voice":"styled"})", no_default, key, "calm");
|
|
check(R"({"input":"x"})", styled, key, "calm");
|
|
for (const std::string & request_key : {key, alias}) {
|
|
check("{\"input\":\"x\",\"voice\":\"styled\",\"options\":{\"" + request_key + "\":\"cheerful\"}}",
|
|
no_default, request_key, "cheerful");
|
|
check("{\"input\":\"x\",\"options\":{\"" + request_key + "\":\"cheerful\"}}",
|
|
styled, request_key, "cheerful");
|
|
const std::string field = request_key == "instruct" ? "instruct" : "instructions";
|
|
check("{\"input\":\"x\",\"voice\":\"styled\",\"" + field + "\":\"cheerful\"}",
|
|
no_default, request_key, "cheerful");
|
|
}
|
|
}
|
|
} catch (const std::exception & error) {
|
|
std::cerr << "voice preset test failed: " << error.what() << "\n";
|
|
return 1;
|
|
}
|
|
std::cout << "server voice preset passed\n";
|
|
return 0;
|
|
}
|