mel_spec 0.4.0

Mel spectrograms aligned to the results from the whisper.cpp, pytorch and librosa reference implementations and suited to streaming audio.
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
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
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153

use crate::mel::{interleave_frames, log_mel_spectrogram, mel, norm_mel};
use crate::quant::quantize;
use crate::stft::Spectrogram;
use crate::vad::{duration_ms_for_n_frames, DetectionSettings, VoiceActivityDetector};
use js_sys::{Object, Reflect, Uint8Array, Uint8ClampedArray};
use ndarray::Array2;
use wasm_bindgen::prelude::*;
use web_sys::Worker;

#[wasm_bindgen]
pub struct SpeechToMel {
    mel: Array2<f64>,
    mel_vad: Array2<f64>,
    fft: Spectrogram,
    vad: VoiceActivityDetector,
    hop_size: usize,
    sampling_rate: f64,
    accumulated_samples: Vec<f32>,
    idx: usize,
}

#[wasm_bindgen]
impl SpeechToMel {
    #[wasm_bindgen]
    pub fn new(fft_size: usize, hop_size: usize, sampling_rate: f64, n_mels: usize) -> Self {
        Self::new_with_settings(
            fft_size,
            hop_size,
            sampling_rate,
            n_mels,
            DetectionSettings {
                min_energy: 1.0,
                min_y: 3,
                min_x: 3,
                min_mel: 0,
            },
        )
    }

    #[wasm_bindgen(js_name = newWithVadSettings)]
    pub fn new_with_vad_settings(
        fft_size: usize,
        hop_size: usize,
        sampling_rate: f64,
        n_mels: usize,
        min_energy: f64,
        min_y: usize,
        min_x: usize,
        min_mel: usize,
    ) -> Self {
        Self::new_with_settings(
            fft_size,
            hop_size,
            sampling_rate,
            n_mels,
            DetectionSettings {
                min_energy,
                min_y,
                min_x,
                min_mel,
            },
        )
    }

    fn new_with_settings(
        fft_size: usize,
        hop_size: usize,
        sampling_rate: f64,
        n_mels: usize,
        settings: DetectionSettings,
    ) -> Self {
        let filters = mel(sampling_rate, fft_size, n_mels, None, None, false, true);
        let filters2 = mel(sampling_rate, fft_size, n_mels / 4, None, None, false, true);
        let stft = Spectrogram::new(fft_size, hop_size);
        let vad = VoiceActivityDetector::new(&settings);
        Self {
            accumulated_samples: Vec::new(),
            mel: filters,
            mel_vad: filters2,
            fft: stft,
            vad,
            sampling_rate,
            hop_size,
            idx: 0,
        }
    }

    #[wasm_bindgen]
    pub fn get(&mut self) -> JsValue {
        let empty = vec![0.0; 0];
        self.add(empty, false)
    }

    #[wasm_bindgen]
    pub fn add(&mut self, data: Vec<f32>, vad: bool) -> JsValue {
        let result = Object::new();
        Reflect::set(&result, &JsValue::from_str("ok"), &JsValue::from(false)).unwrap();
        self.accumulated_samples.extend_from_slice(&data);
        if self.accumulated_samples.len() >= self.hop_size {
            let (samples, rest) = self.accumulated_samples.split_at(self.hop_size);

            Reflect::set(
                &result,
                &JsValue::from_str("len"),
                &JsValue::from(samples.len()),
            )
            .unwrap();

            if let Some(fft) = self.fft.add(&samples.to_vec()) {
                //let frame = norm_mel(&log_mel_spectrogram(&fft, &self.mel));
                let frame = &log_mel_spectrogram(&fft, &self.mel);
                let frame2 = norm_mel(&log_mel_spectrogram(&fft, &self.mel_vad));
                let (quant_frame, range) = quantize(&interleave_frames(&[frame.clone()], false, 0));
                let frame_array = Uint8Array::from(&quant_frame[..]);
                let frame_clamped_array = Uint8ClampedArray::new(&frame_array.buffer());
                Reflect::set(&result, &JsValue::from_str("frame"), &frame_clamped_array).unwrap();
                Reflect::set(&result, &JsValue::from_str("ok"), &JsValue::from(true)).unwrap();
                Reflect::set(
                    &result,
                    &JsValue::from_str("min"),
                    &JsValue::from(range.min),
                )
                .unwrap();
                Reflect::set(
                    &result,
                    &JsValue::from_str("max"),
                    &JsValue::from(range.max),
                )
                .unwrap();
                let ms = duration_ms_for_n_frames(self.hop_size, self.sampling_rate, self.idx);
                Reflect::set(&result, &JsValue::from_str("idx"), &JsValue::from(self.idx)).unwrap();
                Reflect::set(&result, &JsValue::from_str("ms"), &JsValue::from(ms)).unwrap();
                if vad {
                    if let Some(gap) = self.vad.add(&frame2) {
                        Reflect::set(&result, &JsValue::from_str("va"), &JsValue::from(gap))
                            .unwrap();
                    }
                }
            }
            self.accumulated_samples = rest.to_vec();
            self.idx = self.idx.wrapping_add(1);
        }

        JsValue::from(result)
    }
}

/// Run entry point for the main thread.
#[wasm_bindgen]
pub fn startup(path: String) -> Worker {
    let worker_handle = Worker::new(&path).unwrap();
    worker_handle
}