File size: 8,035 Bytes
c6eaeea
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
"""Pick the Space's example clips from material Routyo never trained on, and build two mixed conversations.

    python -I tools/pick_examples.py <routyo package folder>

Synthetic: the test split only (voices held out, everyday sentences from the 10 test-only situations), gate passed.
Real: FLEURS (CC BY 4.0) and VoxPopuli (CC0) test clips for en/fr/es.
Each candidate must be called right by Routyo with the word check; the first that is wins (fixed seed order).
Writes examples/*.wav (16 kHz mono) and examples/examples.json.
"""
import json
import os
import random
import sys

import numpy as np
import soundfile as sf

sys.path.insert(0, sys.argv[1])
from routyo import Routyo, load_audio  # noqa: E402

DATA = os.path.expanduser("~/Documents/Youtube-Ht-Data/data/langid")
WORLD = os.path.join(DATA, "world")
OUT = os.path.join(os.path.dirname(os.path.dirname(os.path.abspath(__file__))), "examples")
os.makedirs(OUT, exist_ok=True)

LOCALE = {"en-us": "American", "en-gb": "British", "fr-fr": "French", "es-es": "Spanish", "es-mx": "Mexican",
          "de-de": "German", "pt-br": "Brazilian"}
voices = json.load(open(os.path.join(WORLD, "voices.json")))
catalogue = {v["id"]: v for v in voices["catalogue"]}
gate = {}
for line in open(os.path.join(WORLD, "gate.jsonl")):
    g = json.loads(line)
    gate[g["id"]] = g["cer"]
made = []
for name in sorted(os.listdir(WORLD)):
    if name.startswith("made_") and "convo" in name:
        for line in open(os.path.join(WORLD, name)):
            m = json.loads(line)
            if m["split"] == "test" and gate.get(m["id"], 1) <= 0.35 and 1.8 <= m["duration"] <= 7:
                made.append(m)
situation = {}
for line in open(os.path.join(WORLD, "convo.jsonl")):
    c = json.loads(line)
    if c["split"] == "test":
        for k in ("ht", "en", "fr", "es"):
            situation[(k, c[k])] = c["situation"]
real = [json.loads(line) for line in open(os.path.join(DATA, "clips.jsonl"))]
real = [r for r in real if r["split"] == "test" and r["source"] in ("fleurs", "voxpopuli_en_accented") and 3 <= r["duration"] <= 9]
print(len(made), "synthetic candidates,", len(real), "real")

r = Routyo()
rng = random.Random(7)


def voice_of(m):
    if m["kind"] == "fatiman_h":
        return "Haitian voice"
    v = catalogue.get(m["voice"], {})
    return f"{LOCALE.get(v.get('locale'), 'foreign')} voice"


def take(pool, lang, label, note, source):
    pool = list(pool)
    rng.shuffle(pool)
    for m in pool:
        x = load_audio(m["path"])
        rates = r.rates(x, words=True)
        if rates[0]["language"] == lang:
            label = label.replace("{voice}", voice_of(m)) if "kind" in m else label
            name = f"{len(picked):02d}_{lang}_{source}.wav"
            sf.write(os.path.join(OUT, name), x, 16000, subtype="PCM_16")
            picked.append({"file": name, "language": lang, "label": label, "note": note, "text": m.get("text", ""),
                           "source": source, "seconds": round(len(x) / 16000, 2), "rate": rates[0]["rate"]})
            print(f"  {label:45} {rates[0]['language']} {rates[0]['rate']:.3f}  {m.get('text', '')[:60]}")
            return m
    print("  none for", label)


def syn(kind, lang, locale=None, gender=None):
    out = []
    for m in made:
        if m["kind"] != kind or m["language"] != lang:
            continue
        if locale and catalogue.get(m["voice"], {}).get("locale") not in locale:
            continue
        out.append({**m, "path": os.path.join(DATA, m["audio"])})
    return out


def rl(source, lang):
    return [{**x, "path": os.path.join(DATA, x["audio"]) if not os.path.isabs(x["audio"]) else x["audio"]}
            for x in real if x["source"] == source and x["language"] == lang]


picked = []
SYN = "synthetic, voice never heard in training"
# Krey貌l
take(syn("fatiman_h", "ht"), "ht", "Krey貌l 路 Haitian voice", SYN, "fatiman")
take(syn("fatiman_c", "ht", {"en-us", "en-gb"}), "ht", "Krey貌l 路 {voice}", SYN, "fatiman-clone")
take(syn("fatiman_c", "ht", {"fr-fr"}), "ht", "Krey貌l 路 {voice}", SYN, "fatiman-clone")
take(syn("fatiman_c", "ht", {"es-mx", "es-es"}), "ht", "Krey貌l 路 {voice}", SYN, "fatiman-clone")
take(syn("phonon", "ht"), "ht", "Krey貌l 路 {voice} (another engine)", SYN, "phonon")
# English
take(syn("fatiman_h", "en"), "en", "English 路 Haitian voice", SYN, "fatiman")
take(syn("phonon", "en"), "en", "English 路 {voice}", SYN, "phonon")
take(syn("fatiman_c", "en", {"de-de", "pt-br"}), "en", "English 路 {voice}", SYN, "fatiman-clone")
take(rl("fleurs", "en"), "en", "English 路 real speaker (FLEURS)", "real recording, FLEURS (CC BY 4.0)", "fleurs")
take(rl("voxpopuli_en_accented", "en"), "en", "English 路 real non-native speaker (VoxPopuli)",
     "real recording, VoxPopuli (CC0)", "voxpopuli")
# French
take(syn("fatiman_h", "fr"), "fr", "Fran莽ais 路 Haitian voice", SYN, "fatiman")
take(syn("phonon", "fr"), "fr", "Fran莽ais 路 {voice}", SYN, "phonon")
take(syn("fatiman_c", "fr", {"en-us", "en-gb"}), "fr", "Fran莽ais 路 {voice}", SYN, "fatiman-clone")
take(rl("fleurs", "fr"), "fr", "Fran莽ais 路 real speaker (FLEURS)", "real recording, FLEURS (CC BY 4.0)", "fleurs")
# Spanish
take(syn("fatiman_h", "es"), "es", "Espa帽ol 路 Haitian voice", SYN, "fatiman")
take(syn("phonon", "es"), "es", "Espa帽ol 路 {voice}", SYN, "phonon")
take(syn("fatiman_c", "es", {"fr-fr"}), "es", "Espa帽ol 路 {voice}", SYN, "fatiman-clone")
take(rl("fleurs", "es"), "es", "Espa帽ol 路 real speaker (FLEURS)", "real recording, FLEURS (CC BY 4.0)", "fleurs")


# Conversations: lines of one test situation, glued with pauses; every line must also be routed right
def right(m):
    return r.rates(load_audio(os.path.join(DATA, m["audio"])), words=True)[0]["language"] == m["language"]


def first_right(c):
    c = list(c)
    rng.shuffle(c)
    return next((m for m in c if right(m)), None)


def glue(lines, name, title):
    gap = np.zeros(int(1.5 * 16000), np.float32)
    parts, script = [], []
    for m in lines:
        x = load_audio(os.path.join(DATA, m["audio"]))
        x = x / (np.abs(x).max() + 1e-6) * 0.8
        parts += [x, gap]
        script.append({"language": m["language"], "text": m["text"], "voice": voice_of(m)})
    y = np.concatenate(parts[:-1]).astype(np.float32)
    sf.write(os.path.join(OUT, name), y, 16000, subtype="PCM_16")
    convos.append({"file": name, "title": title, "lines": script, "seconds": round(len(y) / 16000, 1)})
    print(" ", title, [s["language"] for s in script])


convos = []
by_sit = {}
for m in made:
    s = situation.get((m["language"], m["text"]))
    if s:
        by_sit.setdefault(s, []).append(m)
# 1: many voices, languages alternating, one situation
order = ["ht", "en", "fr", "ht", "es", "en", "ht"]
for sit in sorted(by_sit):
    pool, lines, used = by_sit[sit], [], set()
    for k in order:
        m = first_right(m for m in pool if m["language"] == k and m["text"] not in used)
        if m is None:
            break
        used.add(m["text"])
        lines.append(m)
    if len(lines) == len(order):
        glue(lines, "convo_many_voices.wav", f"Many voices, four languages ({sit})")
        break
# 2: one Haitian voice switching language every sentence
haitian = {}
for m in made:
    if m["kind"] == "fatiman_h":
        haitian.setdefault(m["voice"], set()).add(m["language"])
for voice in sorted(haitian, key=lambda v: -sum(m["voice"] == v for m in made)):
    pool = [m for m in made if m["kind"] == "fatiman_h" and m["voice"] == voice]
    lines = []
    for k in ["ht", "en", "fr", "es", "ht", "fr"]:
        m = first_right(m for m in pool if m["language"] == k and m not in lines)
        if m is None:
            break
        lines.append(m)
    if len(lines) == 6:
        glue(lines, "convo_one_voice.wav", "One Haitian voice switching languages")
        break

json.dump({"examples": picked, "conversations": convos}, open(os.path.join(OUT, "examples.json"), "w"),
          indent=1, ensure_ascii=False)
print(len(picked), "examples,", len(convos), "conversations ->", OUT)