zipvoice_cpp_server/zipvoice_resident_server.cpp

#include "zipvoice_resident_server.hpp"

#include <algorithm>
#include <cmath>
#include <cstdio>
#include <cstring>
#include <fstream>
#include <sstream>

#include "src/EngineWrapper.hpp"
#include "src/fbank.hpp"
#include "src/tokenizer.hpp"
#include "src/vocoder.hpp"
#include "src/zipvoice_engine.hpp"
#include "src/wav_writer.hpp"

// ---------------------------------------------------------------------------
// Simple WAV writer (inlined; avoids external soundfile dependency)
// ---------------------------------------------------------------------------

static bool WriteWav(const std::string& path, const std::vector<float>& samples,
                     int sample_rate, int bits_per_sample = 16) {
    std::ofstream file(path, std::ios::binary);
    if (!file.is_open()) return false;

    int num_samples = (int)samples.size();
    int byte_rate = sample_rate * (bits_per_sample / 8);
    int block_align = bits_per_sample / 8;
    int data_size = num_samples * block_align;

    // RIFF header
    file.write("RIFF", 4);
    uint32_t chunk_size = 36 + data_size;
    file.write(reinterpret_cast<const char*>(&chunk_size), 4);
    file.write("WAVE", 4);

    // fmt subchunk
    file.write("fmt ", 4);
    uint32_t subchunk1_size = 16;
    uint16_t audio_format = 1;      // PCM
    uint16_t num_channels = 1;      // mono
    uint16_t bps = (uint16_t)bits_per_sample;
    file.write(reinterpret_cast<const char*>(&subchunk1_size), 4);
    file.write(reinterpret_cast<const char*>(&audio_format), 2);
    file.write(reinterpret_cast<const char*>(&num_channels), 2);
    uint32_t sr = (uint32_t)sample_rate;
    file.write(reinterpret_cast<const char*>(&sr), 4);
    uint32_t br = (uint32_t)byte_rate;
    file.write(reinterpret_cast<const char*>(&br), 4);
    file.write(reinterpret_cast<const char*>(&block_align), 2);
    file.write(reinterpret_cast<const char*>(&bps), 2);

    // data subchunk
    file.write("data", 4);
    file.write(reinterpret_cast<const char*>(&data_size), 4);

    // Write samples as int16 PCM
    for (float s : samples) {
        float clamped = std::max(-1.0f, std::min(1.0f, s));
        int16_t v = (int16_t)(clamped * 32767.0f);
        file.write(reinterpret_cast<const char*>(&v), sizeof(v));
    }

    file.close();
    return true;
}

// ---------------------------------------------------------------------------
// WAV file helpers (ported from zipvoice.cpp)
// ---------------------------------------------------------------------------

static int ReadWavFile(const std::string& path, std::vector<float>& samples, int& sample_rate) {
    std::ifstream file(path, std::ios::binary);
    if (!file.is_open()) {
        fprintf(stderr, "Failed to open: %s\n", path.c_str());
        return -1;
    }
    char riff[5] = {};
    file.read(riff, 4);
    if (std::strncmp(riff, "RIFF", 4) != 0) {
        fprintf(stderr, "Not a valid WAV: %s\n", path.c_str());
        return -1;
    }
    uint32_t file_size;
    file.read(reinterpret_cast<char*>(&file_size), 4);
    char wave[5] = {};
    file.read(wave, 4);
    if (std::strncmp(wave, "WAVE", 4) != 0) {
        fprintf(stderr, "Not a valid WAV: %s\n", path.c_str());
        return -1;
    }
    int num_channels = 1, bits_per_sample = 16;
    sample_rate = 24000;
    uint32_t data_size = 0;
    while (file.good()) {
        char chunk_id[5] = {};
        file.read(chunk_id, 4);
        uint32_t chunk_size;
        file.read(reinterpret_cast<char*>(&chunk_size), 4);
        if (std::strncmp(chunk_id, "fmt ", 4) == 0) {
            uint16_t fmt, ch, bps; uint32_t sr, br; uint16_t ba;
            file.read(reinterpret_cast<char*>(&fmt), 2);
            file.read(reinterpret_cast<char*>(&ch), 2);
            file.read(reinterpret_cast<char*>(&sr), 4);
            file.read(reinterpret_cast<char*>(&br), 4);
            file.read(reinterpret_cast<char*>(&ba), 2);
            file.read(reinterpret_cast<char*>(&bps), 2);
            num_channels = ch; sample_rate = sr; bits_per_sample = bps;
            if (chunk_size > 16) file.seekg(chunk_size - 16, std::ios::cur);
        } else if (std::strncmp(chunk_id, "data", 4) == 0) {
            data_size = chunk_size;
            break;
        } else {
            file.seekg(chunk_size, std::ios::cur);
        }
    }
    if (data_size == 0) { fprintf(stderr, "No data chunk\n"); return -1; }
    int num_samples = data_size / (bits_per_sample / 8) / num_channels;
    if (bits_per_sample == 16) {
        std::vector<int16_t> raw(num_samples * num_channels);
        file.read(reinterpret_cast<char*>(raw.data()), data_size);
        samples.resize(num_samples);
        for (int i = 0; i < num_samples; ++i)
            samples[i] = raw[i * num_channels] / 32768.0f;
    } else if (bits_per_sample == 32) {
        samples.resize(num_samples * num_channels);
        file.read(reinterpret_cast<char*>(samples.data()), data_size);
        std::vector<float> mono(num_samples);
        for (int i = 0; i < num_samples; ++i) mono[i] = samples[i * num_channels];
        samples = std::move(mono);
    } else {
        fprintf(stderr, "Unsupported bit depth: %d\n", bits_per_sample);
        return -1;
    }
    return 0;
}

static std::vector<float> ResampleLinear(const std::vector<float>& samples,
                                          int orig_sr, int target_sr) {
    if (orig_sr == target_sr) return samples;
    int old_len = (int)samples.size();
    int new_len = std::max(1, (int)std::round((double)old_len * target_sr / orig_sr));
    std::vector<float> result(new_len);
    for (int i = 0; i < new_len; ++i) {
        double pos = (double)i * (old_len - 1) / (new_len - 1);
        int idx = (int)pos;
        double frac = pos - idx;
        if (idx + 1 < old_len)
            result[i] = (float)(samples[idx] * (1.0 - frac) + samples[idx + 1] * frac);
        else
            result[i] = samples[old_len - 1];
    }
    return result;
}

static float ComputeRms(const std::vector<float>& samples) {
    if (samples.empty()) return 0.0f;
    float sum_sq = 0.0f;
    for (float s : samples) sum_sq += s * s;
    return std::sqrt(sum_sq / (float)samples.size());
}

static void RmsNormalize(std::vector<float>& samples, float target_rms) {
    float rms = ComputeRms(samples);
    if (rms < target_rms && rms > 1e-10f) {
        float gain = target_rms / rms;
        for (float& s : samples) s *= gain;
    }
}

// ---------------------------------------------------------------------------
// UTF-8 / Chinese helpers (ported from zipvoice.cpp)
// ---------------------------------------------------------------------------

static bool IsUtf8Lead(unsigned char c) { return (c & 0xC0) != 0x80; }

static std::vector<std::string> SplitUtf8Chars(const std::string& text) {
    std::vector<std::string> chars;
    for (size_t i = 0; i < text.size();) {
        unsigned char c = (unsigned char)text[i];
        size_t len = 1;
        if ((c & 0x80) == 0) len = 1;
        else if ((c & 0xE0) == 0xC0) len = 2;
        else if ((c & 0xF0) == 0xE0) len = 3;
        else if ((c & 0xF8) == 0xF0) len = 4;
        chars.push_back(text.substr(i, len));
        i += len;
    }
    return chars;
}

static bool IsChineseUtf8Char(const std::string& ch) {
    if (ch.size() != 3) return false;
    unsigned char b0 = (unsigned char)ch[0], b1 = (unsigned char)ch[1], b2 = (unsigned char)ch[2];
    uint32_t cp = ((b0 & 0x0F) << 12) | ((b1 & 0x3F) << 6) | (b2 & 0x3F);
    return cp >= 0x4E00 && cp <= 0x9FFF;
}

static bool IsSplitPunct(const std::string& ch) {
    static const char* puncts[] = {
        ".", "!", "?", ";", ",", ":",
        "。", "！", "？", "；", "，", "、", "："
    };
    for (auto* p : puncts) if (ch == p) return true;
    return false;
}

static std::string TrimAsciiSpaces(const std::string& s) {
    size_t start = 0, end = s.size();
    while (start < end && (s[start] == ' ' || s[start] == '\t' || s[start] == '\n' || s[start] == '\r')) start++;
    while (end > start && (s[end-1] == ' ' || s[end-1] == '\t' || s[end-1] == '\n' || s[end-1] == '\r')) end--;
    return s.substr(start, end - start);
}

static std::string JoinUnits(const std::string& left, const std::string& right) {
    if (left.empty()) return TrimAsciiSpaces(right);
    std::string r = TrimAsciiSpaces(right);
    if (r.empty()) return TrimAsciiSpaces(left);
    auto lchars = SplitUtf8Chars(left), rchars = SplitUtf8Chars(r);
    bool zh = (!lchars.empty() && IsChineseUtf8Char(lchars.back())) ||
              (!rchars.empty() && IsChineseUtf8Char(rchars.front()));
    return zh ? (TrimAsciiSpaces(left) + r) : (TrimAsciiSpaces(left) + " " + r);
}

static std::vector<std::string> SplitUnits(const std::string& text) {
    std::vector<std::string> units;
    auto chars = SplitUtf8Chars(TrimAsciiSpaces(text));
    std::string current;
    for (const auto& ch : chars) {
        current += ch;
        if (IsSplitPunct(ch)) {
            std::string t = TrimAsciiSpaces(current);
            if (!t.empty()) units.push_back(t);
            current.clear();
        }
    }
    current = TrimAsciiSpaces(current);
    if (!current.empty()) units.push_back(current);
    if (units.empty() && !text.empty()) units.push_back(TrimAsciiSpaces(text));
    return units;
}

// ---------------------------------------------------------------------------
// Segment estimation / building (ported from zipvoice.cpp BuildSegmentsCpp)
// ---------------------------------------------------------------------------

struct SegmentInfo {
    std::string text;
    int text_tokens = 0;
    int raw_features_len = 0;
    int features_len = 0;
    int generated_frames = 0;
};

static int TokenCount(Tokenizer& tok, const std::string& text) {
    return (int)tok.TextToTokenIds(text).size();
}

static std::vector<std::string> SplitLongUnit(Tokenizer& tok, const std::string& unit,
                                               int max_text_tokens) {
    if (TokenCount(tok, unit) <= max_text_tokens) return {unit};
    std::vector<std::string> chunks;
    if (unit.find(' ') != std::string::npos) {
        std::stringstream ss(unit);
        std::string piece, current;
        while (ss >> piece) {
            std::string candidate = JoinUnits(current, piece);
            if (!current.empty() && TokenCount(tok, candidate) > max_text_tokens) {
                chunks.push_back(current); current = piece;
            } else { current = candidate; }
        }
        if (!current.empty()) chunks.push_back(current);
        return chunks;
    }
    auto chars = SplitUtf8Chars(unit);
    std::string current;
    for (const auto& ch : chars) {
        std::string candidate = current + ch;
        if (!current.empty() && TokenCount(tok, candidate) > max_text_tokens) {
            chunks.push_back(current); current = ch;
        } else { current = candidate; }
    }
    if (!current.empty()) chunks.push_back(current);
    return chunks;
}

static SegmentInfo EstimateSegment(Tokenizer& tok, const std::string& text,
                                    int prompt_frames, int prompt_tokens_len,
                                    float speed, int max_feat_len) {
    SegmentInfo s;
    s.text = text;
    s.text_tokens = TokenCount(tok, text);
    s.raw_features_len = (int)std::ceil((double)prompt_frames / prompt_tokens_len *
                                        (prompt_tokens_len + s.text_tokens) / speed);
    s.features_len = std::min(s.raw_features_len, max_feat_len);
    s.generated_frames = s.features_len - prompt_frames;
    if (s.generated_frames <= 0) s.generated_frames = s.features_len;
    return s;
}

static std::vector<SegmentInfo> BuildSegments(Tokenizer& tok, const std::string& text,
                                               int prompt_frames, int prompt_tokens_len,
                                               float speed, int max_feat_len,
                                               int max_text_tokens, int min_gen_frames,
                                               int max_gen_frames, double max_raw_ratio) {
    auto raw_units = SplitUnits(text);
    std::vector<std::string> units;
    for (const auto& u : raw_units) {
        auto split = SplitLongUnit(tok, u, max_text_tokens);
        units.insert(units.end(), split.begin(), split.end());
    }
    std::vector<SegmentInfo> segments;
    std::string current;
    for (const auto& unit : units) {
        std::string candidate = JoinUnits(current, unit);
        auto cand = EstimateSegment(tok, candidate, prompt_frames, prompt_tokens_len,
                                    speed, max_feat_len);
        bool raw_too_long = cand.raw_features_len > (int)(max_feat_len * max_raw_ratio);
        bool too_long = cand.text_tokens > max_text_tokens ||
                        cand.generated_frames > max_gen_frames || raw_too_long;
        if (!current.empty() && too_long) {
            segments.push_back(EstimateSegment(tok, current, prompt_frames,
                                               prompt_tokens_len, speed, max_feat_len));
            current = unit;
        } else { current = candidate; }
    }
    if (!current.empty())
        segments.push_back(EstimateSegment(tok, current, prompt_frames,
                                           prompt_tokens_len, speed, max_feat_len));
    if (segments.size() >= 2 && segments.back().generated_frames < min_gen_frames) {
        std::string merged = JoinUnits(segments[segments.size()-2].text, segments.back().text);
        auto m = EstimateSegment(tok, merged, prompt_frames, prompt_tokens_len, speed, max_feat_len);
        bool raw_ok = m.raw_features_len <= (int)(max_feat_len * max_raw_ratio);
        if (m.text_tokens <= max_text_tokens && m.generated_frames <= max_gen_frames && raw_ok) {
            segments[segments.size()-2] = m;
            segments.pop_back();
        }
    }
    return segments;
}

// ---------------------------------------------------------------------------
// ZipVoiceResidentServer implementation
// ---------------------------------------------------------------------------

static double GetCurrentTimeMs() {
    struct timeval tv; gettimeofday(&tv, nullptr);
    return tv.tv_sec * 1000.0 + tv.tv_usec / 1000.0;
}

ZipVoiceResidentServer::ZipVoiceResidentServer() = default;
ZipVoiceResidentServer::~ZipVoiceResidentServer() = default;

bool ZipVoiceResidentServer::Init(const std::string& root_dir,
                                   const std::string& model_dir,
                                   const std::string& token_file,
                                   const std::string& vocoder_model,
                                   int max_tokens, int max_feat_len,
                                   int num_step, float default_speed, int seed) {
    root_dir_ = root_dir;
    model_dir_ = model_dir;
    token_file_ = token_file;
    vocoder_model_ = vocoder_model;
    max_tokens_ = max_tokens;
    max_feat_len_ = max_feat_len;
    num_step_ = num_step;
    default_speed_ = default_speed;
    seed_ = seed;

    // 1. Init AX_SYS
    fprintf(stderr, "[zipvoice_daemon] Init AX_SYS...\n");
    if (!InitAxSys()) { fprintf(stderr, "[zipvoice_daemon] AX_SYS_Init FAILED\n"); return false; }
    fprintf(stderr, "[zipvoice_daemon] AX_SYS OK\n");

    // 2. Load tokenizer
    fprintf(stderr, "[zipvoice_daemon] Loading tokenizer...\n");
    if (!LoadTokenizer()) { fprintf(stderr, "[zipvoice_daemon] LoadTokenizer FAILED\n"); return false; }
    fprintf(stderr, "[zipvoice_daemon] Tokenizer OK\n");

    // 3. Load ZipVoice engine (encoder + decoder4)
    fprintf(stderr, "[zipvoice_daemon] Loading engine...\n");
    if (!LoadRuntime()) { fprintf(stderr, "[zipvoice_daemon] LoadRuntime FAILED\n"); return false; }
    fprintf(stderr, "[zipvoice_daemon] Engine OK\n");

    // 4. Load vocoder
    fprintf(stderr, "[zipvoice_daemon] Loading vocoder...\n");
    if (!LoadVocoder()) { fprintf(stderr, "[zipvoice_daemon] LoadVocoder FAILED\n"); return false; }
    fprintf(stderr, "[zipvoice_daemon] Vocoder OK\n");

    // 5. Cache both prompt feature sets
    fprintf(stderr, "[zipvoice_daemon] Caching prompts...\n");
    if (!CachePrompts()) { fprintf(stderr, "[zipvoice_daemon] CachePrompts FAILED\n"); return false; }
    fprintf(stderr, "[zipvoice_daemon] Prompts cached OK\n");

    return true;
}

bool ZipVoiceResidentServer::InitAxSys() {
    int ret = AX_SYS_Init();
    if (ret != 0) {
        fprintf(stderr, "AX_SYS_Init failed: 0x%x\n", ret);
        return false;
    }

    AX_ENGINE_NPU_ATTR_T npu_attr;
    memset(&npu_attr, 0, sizeof(npu_attr));
    npu_attr.eHardMode = static_cast<AX_ENGINE_NPU_MODE_T>(0);
    ret = AX_ENGINE_Init(&npu_attr);
    if (ret != 0) {
        fprintf(stderr, "AX_ENGINE_Init failed: 0x%x\n", ret);
        return false;
    }

    ax_inited_ = true;
    return true;
}

bool ZipVoiceResidentServer::LoadTokenizer() {
    auto tok = std::make_unique<Tokenizer>();
    if (tok->Load(token_file_) != 0) {
        fprintf(stderr, "Tokenizer load failed: %s\n", token_file_.c_str());
        return false;
    }
    tokenizer_ = std::move(tok);
    return true;
}

bool ZipVoiceResidentServer::LoadRuntime() {
    auto eng = std::make_unique<ZipVoiceEngine>();
    if (eng->Init(model_dir_) != 0) {
        fprintf(stderr, "ZipVoiceEngine init failed: %s\n", model_dir_.c_str());
        return false;
    }
    engine_ = std::move(eng);
    return true;
}

bool ZipVoiceResidentServer::LoadVocoder() {
    auto v = std::make_unique<Vocoder>();
    Vocoder::Config vcfg;
    vcfg.model_path = vocoder_model_;
    if (v->Init(vcfg) != 0) {
        fprintf(stderr, "Vocoder init failed: %s\n", vocoder_model_.c_str());
        return false;
    }
    vocoder_ = std::move(v);
    return true;
}

bool ZipVoiceResidentServer::CachePrompts() {
    // Prompt definitions: zh + en
    struct PromptDef {
        std::string key, wav, text;
    };
    std::vector<PromptDef> defs = {
        {"zh", root_dir_ + "/assets/moss_prompts/zh_1_4p5s.wav",
         "不管怎么样我和汤姆还是要感谢贝尔卡金的援手"},
        {"en", root_dir_ + "/assets/moss_prompts/en_4_4p5s.wav",
         "This is almost twice the current industry production level per train."},
    };

    MelFilterBank fbank;
    fbank.Init();

    for (const auto& d : defs) {
        ZipVoicePromptCache cache;
        cache.key = d.key;
        cache.prompt_text = d.text;
        cache.prompt_wav = d.wav;

        // Tokenize
        cache.prompt_tokens = tokenizer_->TextToTokenIds(d.text);

        // Read prompt wav
        std::vector<float> wav;
        int wav_sr = 0;
        if (ReadWavFile(d.wav, wav, wav_sr) != 0) return false;

        // Resample to 24kHz if needed
        std::vector<float> resampled = ResampleLinear(wav, wav_sr, 24000);

        // RMS normalize
        float target_rms = 0.1f;
        RmsNormalize(resampled, target_rms);
        float prompt_rms = ComputeRms(resampled);

        // Extract mel features
        auto features = fbank.Extract(resampled, 24000);
        int T = (int)features.size() / 100;
        std::vector<float> feat_flat(features.size());
        std::copy(features.begin(), features.end(), feat_flat.begin());
        for (auto& f : feat_flat) f *= 0.1f;  // feat_scale

        cache.prompt_features = feat_flat;
        cache.prompt_frames = T;
        cache.prompt_rms = prompt_rms;

        prompts_[d.key] = cache;
    }
    return true;
}

ZipVoiceResponse ZipVoiceResidentServer::Infer(const ZipVoiceRequest& req) {
    ZipVoiceResponse resp;
    resp.id = req.id;

    auto it = prompts_.find(req.prompt);
    if (it == prompts_.end()) {
        resp.ok = false;
        resp.error = "Unknown prompt: " + req.prompt;
        return resp;
    }
    auto& prompt = it->second;

    float speed = req.speed > 0.01f ? req.speed : default_speed_;
    float feat_scale = 0.1f, target_rms = 0.1f;
    int min_gen = 360, max_gen = 620;
    double max_raw_ratio = 1.2;

    double t0 = GetCurrentTimeMs();

    // Build segments
    int max_text_tokens = max_tokens_ - (int)prompt.prompt_tokens.size() - 1;
    if (max_text_tokens <= 0) {
        resp.ok = false;
        resp.error = "prompt tokens leave no room";
        return resp;
    }
    auto segments = BuildSegments(*tokenizer_, req.text, prompt.prompt_frames,
                                   (int)prompt.prompt_tokens.size(), speed,
                                   max_feat_len_, max_text_tokens, min_gen, max_gen, max_raw_ratio);

    // Infer each segment
    std::vector<float> all_audio;
    double model_ms = 0.0;
    for (size_t si = 0; si < segments.size(); ++si) {
        auto& seg = segments[si];

        // Tokenize
        auto text_tokens = tokenizer_->TextToTokenIds(seg.text);

        // Build cat_tokens
        int pad_id = tokenizer_->GetPadId();
        std::vector<int> cat = prompt.prompt_tokens;
        cat.insert(cat.end(), text_tokens.begin(), text_tokens.end());
        cat.push_back(pad_id);
        std::vector<int32_t> cat_padded(max_tokens_, pad_id);
        for (size_t i = 0; i < cat.size() && i < (size_t)max_tokens_; ++i)
            cat_padded[i] = (int32_t)cat[i];

        // Run sample
        ZipVoiceEngine::Timing timing;
        std::vector<float> output_features;
        if (engine_->Sample(cat_padded, (int)prompt.prompt_tokens.size(),
                            (int)text_tokens.size(), prompt.prompt_features,
                            prompt.prompt_frames, speed, 1.0f, seed_ + (int)si,
                            output_features, timing) != 0) {
            resp.ok = false;
            resp.error = "Inference failed on segment " + std::to_string(si);
            return resp;
        }
        model_ms += timing.total_time_sec * 1000.0;

        // Vocoder decode
        std::vector<float> audio;
        if (vocoder_->Decode(output_features, timing.generated_frames, feat_scale, audio) != 0) {
            resp.ok = false;
            resp.error = "Vocoder decode failed on segment " + std::to_string(si);
            return resp;
        }

        // RMS normalize
        RmsNormalize(audio, target_rms);
        if (prompt.prompt_rms < target_rms) {
            float scale = prompt.prompt_rms / target_rms;
            for (float& s : audio) s *= scale;
        }

        // Silence between segments
        if (!all_audio.empty()) {
            std::vector<float> silence(24000 * 140 / 1000, 0.0f);  // 140ms
            all_audio.insert(all_audio.end(), silence.begin(), silence.end());
        }
        all_audio.insert(all_audio.end(), audio.begin(), audio.end());
    }

    // Write WAV
    if (!WriteWav(req.output_wav, all_audio, 24000, 16)) {
        resp.ok = false;
        resp.error = "WriteWav failed: " + req.output_wav;
        return resp;
    }

    double wall_ms = GetCurrentTimeMs() - t0;
    double audio_sec = all_audio.size() / 24000.0;
    resp.ok = true;
    resp.path = req.output_wav;
    resp.wall_sec = wall_ms / 1000.0;
    resp.model_sec = model_ms / 1000.0;
    resp.rtf = audio_sec > 0 ? (model_ms / 1000.0 / audio_sec) : 0.0;
    return resp;
}

bool ZipVoiceResidentServer::CachePrompt(const std::string& key,
                                          const std::string& wav_path,
                                          const std::string& text) {
    MelFilterBank fbank;
    fbank.Init();

    ZipVoicePromptCache cache;
    cache.key = key;
    cache.prompt_text = text;
    cache.prompt_wav = wav_path;
    cache.prompt_tokens = tokenizer_->TextToTokenIds(text);

    std::vector<float> wav;
    int wav_sr = 0;
    if (ReadWavFile(wav_path, wav, wav_sr) != 0) return false;
    std::vector<float> resampled = ResampleLinear(wav, wav_sr, 24000);

    float target_rms = 0.1f;
    RmsNormalize(resampled, target_rms);
    float prompt_rms = ComputeRms(resampled);

    auto features = fbank.Extract(resampled, 24000);
    int T = (int)features.size() / 100;
    std::vector<float> feat_flat(features.size());
    std::copy(features.begin(), features.end(), feat_flat.begin());
    for (auto& f : feat_flat) f *= 0.1f;

    cache.prompt_features = feat_flat;
    cache.prompt_frames = T;
    cache.prompt_rms = prompt_rms;

    prompts_[key] = cache;
    fprintf(stderr, "[zipvoice_daemon] Custom prompt cached: key=%s tokens=%d frames=%d\n",
            key.c_str(), (int)cache.prompt_tokens.size(), T);
    return true;
}