online-transducer-decoder-rknn.h
1.8 KB
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
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
// sherpa-onnx/csrc/rknn/online-transducer-decoder-rknn.h
//
// Copyright (c) 2025 Xiaomi Corporation
#ifndef SHERPA_ONNX_CSRC_RKNN_ONLINE_TRANSDUCER_DECODER_RKNN_H_
#define SHERPA_ONNX_CSRC_RKNN_ONLINE_TRANSDUCER_DECODER_RKNN_H_
#include <vector>
#include "sherpa-onnx/csrc/hypothesis.h"
#include "sherpa-onnx/csrc/macros.h"
namespace sherpa_onnx {
struct OnlineTransducerDecoderResultRknn {
/// Number of frames after subsampling we have decoded so far
int32_t frame_offset = 0;
/// The decoded token IDs so far
std::vector<int64_t> tokens;
/// number of trailing blank frames decoded so far
int32_t num_trailing_blanks = 0;
/// timestamps[i] contains the output frame index where tokens[i] is decoded.
std::vector<int32_t> timestamps;
// used only by greedy_search
std::vector<float> previous_decoder_out;
// used only in modified beam_search
Hypotheses hyps;
// used only by modified_beam_search
std::vector<std::vector<float>> previous_decoder_out2;
};
class OnlineTransducerDecoderRknn {
public:
virtual ~OnlineTransducerDecoderRknn() = default;
/* Return an empty result.
*
* To simplify the decoding code, we add `context_size` blanks
* to the beginning of the decoding result, which will be
* stripped by calling `StripPrecedingBlanks()`.
*/
virtual OnlineTransducerDecoderResultRknn GetEmptyResult() const = 0;
/** Strip blanks added by `GetEmptyResult()`.
*
* @param r It is changed in-place.
*/
virtual void StripLeadingBlanks(
OnlineTransducerDecoderResultRknn * /*r*/) const {}
virtual void Decode(std::vector<float> encoder_out,
OnlineTransducerDecoderResultRknn *result) const = 0;
};
} // namespace sherpa_onnx
#endif // SHERPA_ONNX_CSRC_RKNN_ONLINE_TRANSDUCER_DECODER_RKNN_H_