Function normalize_reference_audio

src/models/omnivoice/audio_tokenizer.cpp:413–447

}

NormalizedReferenceAudio normalize_reference_audio(
    const runtime::AudioBuffer & audio,
    const OmniVoiceAudioTokenizerConfig & config,
    const OmniVoiceReferenceAudioOptions & options) {
    auto mono = to_mono(audio);
    if (audio.sample_rate != config.sample_rate) {
        mono = resample_sinc_hann_mono(mono, audio.sample_rate, config.sample_rate);
    }
    NormalizedReferenceAudio normalized = {};
    normalized.reference_rms = engine::audio::root_mean_square_or_throw(mono);
    if (normalized.reference_rms > 0.0F && normalized.reference_rms < kTargetReferenceRms) {
        const float scale = kTargetReferenceRms / normalized.reference_rms;
        for (float & sample : mono) {
            sample *= scale;
        }
    }
    mono = preprocess_reference_mono(std::move(mono), config.sample_rate, options);

    const int64_t hop_length = checked_positive(config.hop_length, "hop_length");
    const int64_t remainder = static_cast<int64_t>(mono.size()) % hop_length;
    if (remainder > 0) {
        mono.resize(mono.size() - static_cast<size_t>(remainder));
    }
    if (mono.empty()) {
        throw std::runtime_error("OmniVoice reference audio becomes empty after hop-length alignment");
    }

    normalized.frames = static_cast<int64_t>(mono.size()) / hop_length;
    normalized.acoustic_samples_24k = mono;
    auto semantic = resample_sinc_hann_mono(mono, config.sample_rate, config.semantic_sample_rate);
    semantic.insert(semantic.begin(), 160, 0.0F);
    semantic.insert(semantic.end(), 160, 0.0F);
    normalized.semantic_samples_16k_padded = std::move(semantic);
    return normalized;
}

int64_t conv1d_output_length(int64_t input, int64_t kernel, int stride, int padding, int dilation) {
    if (input <= 0 || kernel <= 0 || stride <= 0 || dilation <= 0) {