cpp/src/inflect_tts.h

#pragma once

// InflectTTS: two-stage (encoder/decoder AXMODEL) VITS TTS pipeline.
//
// Scope note: the eSpeak text frontend is NOT part of the C++ SDK — the
// caller supplies phoneme token ids (produced by the Python SDK frontend or
// a board-side eSpeak-ng integration; see README). The full duration /
// alignment / expansion / noise / chunking host chain IS implemented here.

#include <cstdint>
#include <string>
#include <vector>

#include "ax_runner.h"

class InflectTTS {
public:
    InflectTTS(const std::string& encoder_path, const std::string& decoder_path);

    // token_ids: raw phoneme ids (NOT interspersed; blank 0 is inserted here,
    // mirroring commons.intersperse with add_blank=true).
    // speed in [0.5, 2.0], variation in [0.0, 1.0].
    // Returns 24 kHz mono float32 samples in [-1, 1].
    std::vector<float> synthesize_tokens(const std::vector<int64_t>& token_ids,
                                         float speed = 1.0f,
                                         float variation = 0.667f,
                                         uint64_t seed = 0);

private:
    AxRunner encoder_;
    AxRunner decoder_;

    // One Tp=512 decoder chunk: z_chunk channel-major [192 * kDecoderTp]
    // (zero right-padded) -> wav [kDecoderTp * 256].
    std::vector<float> run_decoder_chunk(const std::vector<float>& z_chunk);
};