offline-tts.h
3.8 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
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
// sherpa-onnx/csrc/offline-tts.h
//
// Copyright (c) 2023 Xiaomi Corporation
#ifndef SHERPA_ONNX_CSRC_OFFLINE_TTS_H_
#define SHERPA_ONNX_CSRC_OFFLINE_TTS_H_
#include <cstdint>
#include <functional>
#include <memory>
#include <string>
#include <vector>
#include "sherpa-onnx/csrc/offline-tts-model-config.h"
#include "sherpa-onnx/csrc/parse-options.h"
namespace sherpa_onnx {
struct OfflineTtsConfig {
OfflineTtsModelConfig model;
// If not empty, it contains a list of rule FST filenames.
// Filenames are separated by a comma.
// Example value: rule1.fst,rule2,fst,rule3.fst
//
// If there are multiple rules, they are applied from left to right.
std::string rule_fsts;
// If there are multiple FST archives, they are applied from left to right.
std::string rule_fars;
// Maximum number of sentences that we process at a time.
// This is to avoid OOM for very long input text.
// If you set it to -1, then we process all sentences in a single batch.
int32_t max_num_sentences = 1;
// A silence interval contains audio samples with value close to 0.
//
// the duration of the new interval is old_duration * silence_scale.
float silence_scale = 0.2;
OfflineTtsConfig() = default;
OfflineTtsConfig(const OfflineTtsModelConfig &model,
const std::string &rule_fsts, const std::string &rule_fars,
int32_t max_num_sentences, float silence_scale)
: model(model),
rule_fsts(rule_fsts),
rule_fars(rule_fars),
max_num_sentences(max_num_sentences),
silence_scale(silence_scale) {}
void Register(ParseOptions *po);
bool Validate() const;
std::string ToString() const;
};
struct GeneratedAudio {
std::vector<float> samples;
int32_t sample_rate;
// Silence means pause here.
// If scale > 1, then it increases the duration of a pause
// If scale < 1, then it reduces the duration of a pause
GeneratedAudio ScaleSilence(float scale) const;
};
class OfflineTtsImpl;
// If the callback returns 0, then it stops generating
// if the callback returns 1, then it keeps generating
using GeneratedAudioCallback = std::function<int32_t(
const float * /*samples*/, int32_t /*n*/, float /*progress*/)>;
class OfflineTts {
public:
~OfflineTts();
explicit OfflineTts(const OfflineTtsConfig &config);
template <typename Manager>
OfflineTts(Manager *mgr, const OfflineTtsConfig &config);
// @param text A string containing words separated by spaces
// @param sid Speaker ID. Used only for multi-speaker models, e.g., models
// trained using the VCTK dataset. It is not used for
// single-speaker models, e.g., models trained using the ljspeech
// dataset.
// @param speed The speed for the generated speech. E.g., 2 means 2x faster.
// @param callback If not NULL, it is called whenever config.max_num_sentences
// sentences have been processed. Note that the passed
// pointer `samples` for the callback might be invalidated
// after the callback is returned, so the caller should not
// keep a reference to it. The caller can copy the data if
// he/she wants to access the samples after the callback
// returns. The callback is called in the current thread.
GeneratedAudio Generate(const std::string &text, int64_t sid = 0,
float speed = 1.0,
GeneratedAudioCallback callback = nullptr) const;
// Return the sample rate of the generated audio
int32_t SampleRate() const;
// Number of supported speakers.
// If it supports only a single speaker, then it return 0 or 1.
int32_t NumSpeakers() const;
private:
std::unique_ptr<OfflineTtsImpl> impl_;
};
} // namespace sherpa_onnx
#endif // SHERPA_ONNX_CSRC_OFFLINE_TTS_H_