voxr 0.7.0

Audio and speech: codecs, DSP, G2P, voices, and TTS/ASR pipelines over boostr models
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
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342

//! misaki `G2P.preprocess` (markdown-link features) and the feature pass of
//! `G2P.tokenize`, with spaCy's `Alignment.from_strings`.
//!
//! A link `[text](feature)` keeps `text` in the sentence and attaches
//! `feature` to it: an integer or `±0.5` stress, `/phonemes/`, or
//! `#flags#`. `tokenize` finds the spaCy tokens of the link through a
//! character alignment of the whitespace-split words against spaCy's tokens.

use super::chars::is_py_space;
use super::token::MToken;
use crate::error::{Error, Result};

/// One link feature, as misaki's `preprocess` converts it.
#[derive(Debug, Clone, PartialEq)]
pub(crate) enum Feature {
    /// An integer or `±0.5`: `MToken._.stress`.
    Stress(f64),
    /// `/phonemes/`, stored as `/` + the text with trailing `/` removed.
    Phonemes(String),
    /// `#flags#`, stored as `#` + the text with trailing `#` removed.
    NumFlags(String),
}

/// Output of [`preprocess`].
#[derive(Debug, Clone, PartialEq)]
pub(crate) struct Preprocessed {
    /// The text with every link replaced by its link text.
    pub text: String,
    /// Whitespace-split words, one per link text.
    pub words: Vec<String>,
    /// Feature per index into `words`, ascending.
    pub features: Vec<(usize, Feature)>,
}

/// misaki `is_digit`: `re.match(r'^[0-9]+$', text)`, where `$` also
/// matches before one final newline.
fn is_digit(text: &str) -> bool {
    let body = text.strip_suffix('\n').unwrap_or(text);
    !body.is_empty() && body.bytes().all(|b| b.is_ascii_digit())
}

/// misaki's conversion of a link target; `None` when it carries no feature.
fn feature(f: &str) -> Option<Feature> {
    let unsigned = f.strip_prefix(['-', '+']).unwrap_or(f);
    if is_digit(unsigned) {
        // `int(f)`: sign, digits, and the newline `int` ignores.
        let digits: f64 = unsigned.trim_end_matches('\n').parse().ok()?;
        let negative = f.starts_with('-') && digits != 0.0;
        return Some(Feature::Stress(if negative { -digits } else { digits }));
    }
    match f {
        "0.5" | "+0.5" => return Some(Feature::Stress(0.5)),
        "-0.5" => return Some(Feature::Stress(-0.5)),
        _ => {}
    }
    let edged = |c: char| f.chars().count() > 1 && f.starts_with(c) && f.ends_with(c);
    if edged('/') {
        Some(Feature::Phonemes(format!(
            "/{}",
            f[1..].trim_end_matches('/')
        )))
    } else if edged('#') {
        Some(Feature::NumFlags(format!(
            "#{}",
            f[1..].trim_end_matches('#')
        )))
    } else {
        None
    }
}

/// The next `\[([^\]]+)\]\(([^\)]*)\)` match at or after byte `from`:
/// `(start, end, link text, target)`.
fn next_link(text: &str, from: usize) -> Option<(usize, usize, &str, &str)> {
    let mut search = from;
    while let Some(rel) = text[search..].find('[') {
        let open = search + rel;
        let inner = open + 1;
        let close = text[inner..].find(']').map(|i| inner + i);
        if let Some(close) = close.filter(|&c| c > inner)
            && text[close + 1..].starts_with('(')
        {
            let target = close + 2;
            if let Some(end) = text[target..].find(')').map(|i| target + i) {
                return Some((open, end + 1, &text[inner..close], &text[target..end]));
            }
        }
        search = inner;
    }
    None
}

/// Python `str.split()`.
fn py_split(s: &str) -> impl Iterator<Item = &str> {
    s.split(is_py_space).filter(|w| !w.is_empty())
}

/// misaki `G2P.preprocess(text)`.
pub(crate) fn preprocess(text: &str) -> Preprocessed {
    let text = text.trim_start_matches(is_py_space);
    let mut out = Preprocessed {
        text: String::with_capacity(text.len()),
        words: Vec::new(),
        features: Vec::new(),
    };
    let mut last_end = 0;
    while let Some((start, end, link, target)) = next_link(text, last_end) {
        out.text.push_str(&text[last_end..start]);
        out.words
            .extend(py_split(&text[last_end..start]).map(str::to_string));
        if let Some(f) = feature(target) {
            out.features.push((out.words.len(), f));
        }
        out.text.push_str(link);
        out.words.push(link.to_string());
        last_end = end;
    }
    if last_end < text.len() {
        out.text.push_str(&text[last_end..]);
        out.words
            .extend(py_split(&text[last_end..]).map(str::to_string));
    }
    out
}

fn misaligned(a: &[String], b: &[String]) -> Error {
    let head = |v: &[String]| v.iter().take(10).cloned().collect::<Vec<_>>();
    Error::DataError {
        reason: format!(
            "misaki cannot align words {:?} with spaCy tokens {:?}",
            head(a),
            head(b)
        ),
    }
}

/// spaCy `get_alignments(A, B)[1]` (`y2x`) flattened: the sorted `A`
/// indices of each `B` token, concatenated, as `Ragged.data` holds them.
pub(crate) fn y2x_data(a: &[String], b: &[String]) -> Result<Vec<usize>> {
    let owners = |v: &[String]| -> Vec<usize> {
        v.iter()
            .enumerate()
            .flat_map(|(i, x)| std::iter::repeat_n(i, x.to_lowercase().chars().count()))
            .collect()
    };
    let (char_a, char_b) = (owners(a), owners(b));
    let str_a: Vec<char> = a.concat().to_lowercase().chars().collect();
    let str_b: Vec<char> = b.concat().to_lowercase().chars().collect();
    let squeeze = |s: &[char]| s.iter().filter(|c| !is_py_space(**c)).collect::<String>();
    if squeeze(&str_a) != squeeze(&str_b)
        || str_a.len() != char_a.len()
        || str_b.len() != char_b.len()
    {
        return Err(misaligned(a, b));
    }
    let a_chars: Vec<usize> = a.iter().map(|x| x.chars().count()).collect();
    let b_chars: Vec<usize> = b.iter().map(|x| x.chars().count()).collect();
    let (mut ia, mut ib) = (0usize, 0usize);
    let mut prev_b = usize::MAX;
    let mut b2a: Vec<Vec<usize>> = Vec::new();
    while ia < str_a.len() && ib < str_b.len() {
        let (ta, tb) = (char_a[ia], char_b[ib]);
        if prev_b != tb {
            b2a.push(Vec::new());
        }
        let starts_a = ia == 0 || char_a[ia - 1] < ta;
        let starts_b = ib == 0 || char_b[ib - 1] < tb;
        let current = b2a.last_mut().ok_or_else(|| misaligned(a, b))?;
        if a[ta] == b[tb] && starts_a && starts_b {
            current.push(ta);
            ia += a_chars[ta];
            ib += b_chars[tb];
        } else if str_a[ia] == str_b[ib] {
            current.push(ta);
            ia += 1;
            ib += 1;
        } else if is_py_space(str_a[ia]) {
            ia += 1;
        } else if is_py_space(str_b[ib]) {
            ib += 1;
        } else {
            return Err(misaligned(a, b));
        }
        prev_b = tb;
    }
    Ok(b2a
        .into_iter()
        .flat_map(|mut xs| {
            xs.sort_unstable();
            xs.dedup();
            xs
        })
        .collect())
}

/// The feature pass of misaki `G2P.tokenize`: applies `features` to the
/// spaCy tokens `tokens` through the alignment with `words`.
///
/// Like misaki, it looks up each feature's word index in the flattened
/// `y2x` data and uses the position found there as the token index.
pub(crate) fn apply_features(
    tokens: &mut [MToken],
    words: &[String],
    features: &[(usize, Feature)],
) -> Result<()> {
    if features.is_empty() {
        return Ok(());
    }
    let texts: Vec<String> = tokens.iter().map(|t| t.text.clone()).collect();
    let data = y2x_data(words, &texts)?;
    for (k, f) in features {
        let hits = data
            .iter()
            .enumerate()
            .filter(|(_, x)| *x == k)
            .map(|(j, _)| j);
        for (i, j) in hits.enumerate() {
            let Some(tk) = tokens.get_mut(j) else {
                continue;
            };
            match f {
                Feature::Stress(s) => tk.stress = Some(*s),
                Feature::Phonemes(p) => {
                    tk.is_head = i == 0;
                    tk.phonemes = Some(if i == 0 {
                        p.trim_start_matches('/').to_string()
                    } else {
                        String::new()
                    });
                    tk.rating = Some(5);
                }
                Feature::NumFlags(n) => tk.num_flags = n.trim_start_matches('#').to_string(),
            }
        }
    }
    Ok(())
}

#[cfg(test)]
mod tests {
    use super::*;

    fn words(v: &[&str]) -> Vec<String> {
        v.iter().map(|s| s.to_string()).collect()
    }

    // Expected values from misaki 0.9.4 `G2P.preprocess`.
    #[test]
    fn preprocess_extracts_links() {
        let p = preprocess("  [Kokoro](/kˈOkəɹO/) is great.");
        assert_eq!(p.text, "Kokoro is great.");
        assert_eq!(p.words, words(&["Kokoro", "is", "great."]));
        assert_eq!(p.features, [(0, Feature::Phonemes("/kˈOkəɹO".into()))]);

        let p = preprocess(
            "[Hello world](/həlˈO/) and [read](-1) [it](+2) [a](0.5) [b](-0.5) [c](#&#) [d](x)",
        );
        assert_eq!(p.text, "Hello world and read it a b c d");
        assert_eq!(
            p.words,
            words(&["Hello world", "and", "read", "it", "a", "b", "c", "d"])
        );
        assert_eq!(
            p.features,
            [
                (0, Feature::Phonemes("/həlˈO".into())),
                (2, Feature::Stress(-1.0)),
                (3, Feature::Stress(2.0)),
                (4, Feature::Stress(0.5)),
                (5, Feature::Stress(-0.5)),
                (6, Feature::NumFlags("#&".into())),
            ]
        );

        let p = preprocess("x [a]() y [5](//ab//) [](z) [[q]](1)");
        assert_eq!(p.text, "x a y 5 [](z) [[q]](1)");
        assert_eq!(p.words, words(&["x", "a", "y", "5", "[](z)", "[[q]](1)"]));
        assert_eq!(p.features, [(3, Feature::Phonemes("//ab".into()))]);

        let p = preprocess("[a](-0) [b](3\n) [c\nd](/x/)");
        assert_eq!(p.text, "a b c\nd");
        assert_eq!(p.words, words(&["a", "b", "c\nd"]));
        assert_eq!(
            p.features,
            [
                (0, Feature::Stress(0.0)),
                (1, Feature::Stress(3.0)),
                (2, Feature::Phonemes("/x".into()))
            ]
        );
    }

    #[test]
    fn preprocess_without_links_is_lstrip() {
        let p = preprocess(" \t Plain text here. ");
        assert_eq!(p.text, "Plain text here. ");
        assert!(p.features.is_empty());
    }

    #[test]
    fn digit_targets_follow_python_int() {
        assert_eq!(feature("007"), Some(Feature::Stress(7.0)));
        assert_eq!(feature("3\n"), Some(Feature::Stress(3.0)));
        assert_eq!(feature("-"), None);
        assert_eq!(feature("1.5"), None);
        assert_eq!(feature("/"), None);
    }

    #[test]
    fn alignment_matches_spacy() {
        // `spacy.training.Alignment.from_strings(A, B).y2x.data`.
        let a = words(&["Say", "don't", "now"]);
        let b = words(&["Say", "do", "n't", "now"]);
        assert_eq!(y2x_data(&a, &b).ok(), Some(vec![0, 1, 1, 2]));
        let a = words(&["New York", "rocks"]);
        let b = words(&["New", "York", "rocks"]);
        assert_eq!(y2x_data(&a, &b).ok(), Some(vec![0, 0, 1]));
        let a = words(&["a", "b"]);
        let b = words(&["a", "\n\n", "b"]);
        assert_eq!(y2x_data(&a, &b).ok(), Some(vec![0, 1]));
        let b = words(&["a", "c"]);
        assert!(y2x_data(&a, &b).is_err());
    }

    #[test]
    fn features_reach_the_link_tokens() {
        let mut tokens = vec![
            MToken::new("New", "NNP", " "),
            MToken::new("York", "NNP", " "),
            MToken::new("rocks", "NNS", ""),
        ];
        let p = preprocess("[New York](/nuː jɔːk/) rocks");
        apply_features(&mut tokens, &p.words, &p.features)
            .unwrap_or_else(|e| panic!("apply_features: {e}"));
        assert_eq!(tokens[0].phonemes.as_deref(), Some("nuː jɔːk"));
        assert!(tokens[0].is_head);
        assert_eq!(tokens[1].phonemes.as_deref(), Some(""));
        assert!(!tokens[1].is_head);
        assert_eq!((tokens[0].rating, tokens[1].rating), (Some(5), Some(5)));
        assert_eq!(tokens[2].phonemes, None);
    }
}