RL10x's picture
Add inference code (Python + C++ engine + web page)
ab1b068 verified
Raw History Blame Contribute Delete
2.54 kB
// Hindi tiny-voca TTS engine on ggml: phoneme ids -> waveform (24 kHz). Port of tiny_voca.model.TinyVoca.synthesize
// (voiced_vowels=True, soft octave limiter) and the source-filter synthesizer (synth.h).
#pragma once
#include <cstdint>
#include <memory>
#include <stdexcept>
#include <string>
#include <vector>
struct TtsOptions {
int style = 0; // index into TtsModel::styles()
float speed = 1.0f; // 0.5 .. 2
float octave_fix = 1.6f; // soft pitch limiter (0 = off), infer.OCTAVE_FIX
uint64_t seed = 0; // noise seed
// optional external noise, for checking against PyTorch: synthesizer models T*hop samples; flow models the starting
// point x0 of the flow, T*n_mels values with the temperature already applied
const float* noise = nullptr;
int flow_steps = 0; // flow models: midpoint steps (0 = the GGUF's default, 16)
float temperature = -1.0f; // flow models: starting-noise scale (< 0 = the GGUF's default, 1.0)
float guidance = -1.0f; // flow models: classifier-free guidance strength (< 0 = the GGUF's default)
// prosody-flow models, for checking against PyTorch: the duration flow's start [N] and the pitch/energy flow's
// start [T * 2], temperature applied
const float* duration_noise = nullptr;
const float* pitch_energy_noise = nullptr;
};
// Intermediates, filled when a pointer is passed (for checking against scripts/tts_hi_reference.py and
// scripts/tts_hi_flow_reference.py).
struct TtsDebug {
std::vector<float> encoder, duration_raw, prosody_hidden, prosody_raw, prosody, acoustic, f0, harmonic_head, noise_head;
std::vector<float> flow_cond, mel; // flow models: [T][259] conditioning and [T][n_mels] flow output
std::vector<int32_t> durations;
int frames = 0;
};
class TtsModel {
public:
static std::unique_ptr<TtsModel> load(const std::string& gguf_path, int threads);
virtual ~TtsModel() = default;
virtual std::vector<float> synthesize(const std::vector<int32_t>& ids, const std::vector<int32_t>& stress,
const TtsOptions& opt, TtsDebug* dbg = nullptr) = 0;
virtual const std::vector<std::string>& styles() const = 0;
virtual const std::vector<std::string>& symbols() const = 0;
virtual int sample_rate() const = 0;
virtual int max_tokens() const = 0;
virtual size_t weight_bytes() const = 0;
};
class UtteranceTooLong : public std::runtime_error {
public:
using std::runtime_error::runtime_error;
};