transcribe-cpp-sys 0.0.6

Native FFI bindings for transcribe.cpp, a C/C++ speech-to-text library built on ggml
Documentation 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47

// arch/canary_qwen/encoder.h - perception encoder graph builder.
//
// FastConformer encoder (32 blocks) followed by perception projection
// (Linear(1024, 2048) + bias). The conformer body is byte-for-byte
// identical to canary-1b-flash; the projection is canary_qwen-specific.

#pragma once

#include "ggml.h"

struct ggml_context;
struct ggml_tensor;
struct ggml_cgraph;

namespace transcribe::canary_qwen {

struct CanaryQwenHParams;
struct CanaryQwenWeights;

struct EncoderDumps {
    ggml_tensor * mel_in          = nullptr;
    ggml_tensor * pre_encode_out  = nullptr;
    ggml_tensor * pos_emb         = nullptr;
    ggml_tensor * block0_out      = nullptr;
    ggml_tensor * block_mid_out   = nullptr;   // block n_layers/2
    ggml_tensor * block_last_out  = nullptr;   // block n_layers-1
    ggml_tensor * final_out       = nullptr;   // post-encoder, pre-projection
    ggml_tensor * perception_out  = nullptr;   // post Linear(1024, 2048)
};

struct EncoderBuild {
    ggml_tensor * mel_in     = nullptr;
    ggml_tensor * pos_emb_in = nullptr;
    ggml_tensor * out        = nullptr;   // perception_out
    EncoderDumps  dumps {};
    ggml_cgraph * graph      = nullptr;
};

EncoderBuild build_encoder_graph(ggml_context *             ctx,
                                 const CanaryQwenWeights &  weights,
                                 const CanaryQwenHParams &  hp,
                                 int                        n_mel_frames,
                                 ggml_type                  kv_type      = GGML_TYPE_COUNT,
                                 bool                       use_flash    = false,
                                 const char *               backend_name = "");

} // namespace transcribe::canary_qwen