voice_audio_multi/include/engine/models/owsm_ctc/model.h
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
Initial import: audio.cpp with Orange Pi documentation
2026-11-04 19:05:22 +01:00

99 lines
3.5 KiB
C++

#pragma once
#include "engine/framework/assets/resource_bundle.h"
#include "engine/framework/audio/mel_spectrogram_frontend.h"
#include "engine/framework/core/backend_weight_store.h"
#include "engine/framework/core/execution_context.h"
#include "engine/framework/modules/attention/types.h"
#include "engine/framework/modules/conv_modules.h"
#include "engine/framework/modules/conformer_modules.h"
#include "engine/framework/modules/ebranchformer_modules.h"
#include "engine/framework/modules/feed_forward_modules.h"
#include "engine/framework/modules/norm_modules.h"
#include "engine/framework/modules/streaming_conv_modules.h"
#include "engine/framework/modules/transformers/transformer_blocks.h"
#include "engine/framework/runtime/model.h"
#include "engine/framework/tokenizers/sentencepiece.h"
#include <filesystem>
#include <memory>
#include <string>
#include <vector>
namespace engine::models::owsm_ctc {
struct OWSMCTCV4Config {
std::string variant;
int64_t prompt_hidden_size = 0;
int64_t prompt_num_heads = 0;
int64_t prompt_layers = 0;
int64_t prompt_intermediate_size = 0;
std::vector<int64_t> interctc_layers;
std::vector<int64_t> cross_attention_layers;
int64_t hidden_size = 0;
int64_t num_heads = 0;
int64_t encoder_layers = 0;
int64_t intermediate_size = 0;
int64_t vocabulary_size = 50002;
int64_t max_audio_samples = 480000;
int64_t frontend_frames = 3001;
int64_t encoder_frames = 374;
int32_t blank_id = 0;
};
struct OWSMCTCV4Assets {
assets::ResourceBundle resources;
std::shared_ptr<const assets::TensorSource> source;
std::vector<tokenizers::SentencePiecePiece> sentencepiece;
std::shared_ptr<const audio::MelSpectrogramFrontend> frontend;
std::vector<float> feature_mean;
std::vector<float> feature_std;
OWSMCTCV4Config config;
int32_t token_id(const std::string & token) const;
std::string decode_visible(const std::vector<int32_t> & ids) const;
};
struct OWSMCTCV4Weights {
std::unique_ptr<core::BackendWeightStore> store;
modules::EspnetConv2dSubsampling8Weights subsampling;
std::vector<modules::EBranchformerBlockWeights> encoder;
modules::NormWeights encoder_norm;
core::TensorValue embedding;
modules::LinearWeights output;
core::TensorValue half_scale;
core::TensorValue embedding_scale;
std::vector<modules::TransformerEncoderBlockWeights> prompt_encoder;
modules::NormWeights prompt_norm;
modules::LinearWeights prefix_projection;
modules::LinearWeights prompt_projection;
modules::LinearWeights ctc_conditioning;
core::TensorValue prompt_scale;
};
std::shared_ptr<const OWSMCTCV4Assets> load_owsm_ctc_assets(const std::filesystem::path & path);
std::unique_ptr<OWSMCTCV4Weights> load_owsm_ctc_weights(
const OWSMCTCV4Assets & assets,
core::ExecutionContext & execution,
assets::TensorStorageType type);
class OWSMCTCV4EBranchformerRuntime {
public:
OWSMCTCV4EBranchformerRuntime(const OWSMCTCV4Assets & assets, const OWSMCTCV4Weights & weights,
core::ExecutionContext & execution);
~OWSMCTCV4EBranchformerRuntime();
std::vector<int32_t> frame_tokens(const std::vector<float> & samples,
int32_t language, int32_t task);
private:
struct Graphs;
const OWSMCTCV4Assets & assets_;
const OWSMCTCV4Weights & weights_;
core::ExecutionContext & execution_;
std::unique_ptr<Graphs> graphs_;
};
std::shared_ptr<runtime::IVoiceModelLoader> make_owsm_ctc_loader();
} // namespace engine::models::owsm_ctc