File size: 30,960 Bytes
d127cdf
 
a0c8e96
e1e3b69
720f9d4
d127cdf
87e6aa5
 
 
 
 
 
 
346d2ba
 
87e6aa5
 
 
d127cdf
 
 
 
 
 
 
 
 
 
 
 
 
720f9d4
 
e1e3b69
720f9d4
 
 
e1e3b69
 
 
 
d127cdf
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
65442de
d127cdf
 
 
 
 
 
 
65442de
 
 
d127cdf
 
 
 
720f9d4
 
87e6aa5
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
d127cdf
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
346d2ba
 
87e6aa5
d127cdf
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
5a70a4f
 
 
 
44574cf
5a70a4f
346d2ba
 
 
 
 
 
 
 
 
 
e1e3b69
346d2ba
e1e3b69
 
 
346d2ba
 
e1e3b69
2ab99a5
 
 
 
 
 
346d2ba
 
 
 
 
2ab99a5
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
44574cf
5a70a4f
346d2ba
5a70a4f
44574cf
996d665
d127cdf
 
 
 
 
 
 
 
 
 
 
 
 
44574cf
fa98881
44574cf
996d665
 
d4cfcec
44574cf
996d665
5a70a4f
e1e3b69
44574cf
 
428e952
44574cf
 
d4cfcec
 
5a70a4f
e1e3b69
346d2ba
e1e3b69
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
346d2ba
 
e1e3b69
 
 
 
346d2ba
e1e3b69
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
996d665
e1e3b69
 
 
 
 
996d665
e1e3b69
 
 
 
 
346d2ba
e1e3b69
 
 
 
 
 
 
 
 
 
 
 
 
 
 
28ff1d1
 
 
 
 
 
2ab99a5
 
 
 
 
 
7bcd2b0
 
2ab99a5
 
 
 
 
 
 
 
 
 
 
 
 
28ff1d1
 
 
 
 
 
 
 
2ab99a5
28ff1d1
 
2ab99a5
28ff1d1
2ab99a5
 
28ff1d1
e1e3b69
d127cdf
 
 
 
 
 
996d665
 
 
e1e3b69
996d665
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
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
from __future__ import annotations

import logging
import mimetypes
import re
import sys
import threading
import time
from dataclasses import dataclass
from fractions import Fraction
from pathlib import Path

import gradio as gr
import numpy as np
import torch
from audiotsm import wsola
from audiotsm.io.array import ArrayReader, ArrayWriter
from scipy.signal import resample_poly

try:
    import spaces
except ImportError:  # Local smoke tests do not require the ZeroGPU shim.
    class _Spaces:
        @staticmethod
        def GPU(*args, **kwargs):
            def decorate(function):
                return function
            return decorate
    spaces = _Spaces()


ROOT = Path(__file__).resolve().parent
RUNTIME = ROOT / "runtime"
WEB_RUNTIME = ROOT / "web_runtime"
sys.path.insert(0, str(RUNTIME))
sys.path.insert(0, str(ROOT))

# Windows does not register .mjs by default. Gradio's static-file response
# would otherwise serve ONNX Runtime's module as text/plain.
mimetypes.add_type("text/javascript", ".mjs")
gr.set_static_paths(paths=[WEB_RUNTIME])

import commons
import utils
from inflect_vits_frontend import run_vits_frontend
from models import SynthesizerTrn
from text import cleaned_text_to_sequence
from text.symbols import symbols


@dataclass(frozen=True)
class ModelSpec:
    label: str
    short_label: str
    params: str
    checkpoint: Path
    config: Path
    revision: str


SPECS = {
    "Inflect Micro v2": ModelSpec(
        label="Inflect Micro v2",
        short_label="Micro",
        params="9.36M",
        checkpoint=ROOT / "models" / "micro" / "model.pth",
        config=ROOT / "models" / "micro" / "config.json",
        revision="3eede065",
    ),
    "Inflect Nano v2": ModelSpec(
        label="Inflect Nano v2",
        short_label="Nano",
        params="3.96M",
        checkpoint=ROOT / "models" / "nano" / "model.pth",
        config=ROOT / "models" / "nano" / "config.json",
        revision="bfca4684",
    ),
}


def split_text(text: str, limit: int = 280) -> list[str]:
    normalized = " ".join(text.split())
    sentences = [
        part.strip()
        for part in re.split(r"(?<=[.!?;:])\s+", normalized)
        if part.strip()
    ]
    chunks: list[str] = []
    for sentence in sentences or [normalized]:
        while len(sentence) > limit:
            search = sentence[: limit + 1]
            punctuation = max(search.rfind(mark) for mark in (",", ";", ":"))
            split_at = (
                punctuation + 1
                if punctuation >= limit // 2
                else sentence.rfind(" ", 0, limit + 1)
            )
            if split_at < limit // 2:
                split_at = limit
            chunks.append(sentence[:split_at].strip())
            sentence = sentence[split_at:].strip()
        if sentence:
            chunks.append(sentence)
    return chunks


def boundary_pause_seconds(chunk: str) -> float:
    ending = chunk.rstrip()[-1:] if chunk.strip() else ""
    return {
        "?": 0.28,
        "!": 0.24,
        ".": 0.22,
        ";": 0.16,
        ":": 0.13,
        ",": 0.09,
    }.get(ending, 0.08)


def edge_fade(waveform: np.ndarray, sample_rate: int, milliseconds: float = 5.0) -> np.ndarray:
    frames = min(round(sample_rate * milliseconds / 1000.0), waveform.size // 2)
    if frames <= 0:
        return waveform
    output = waveform.copy()
    ramp = np.linspace(0.0, 1.0, frames, endpoint=True, dtype=np.float32)
    output[:frames] *= ramp
    output[-frames:] *= ramp[::-1]
    return output


def pcm16(waveform: np.ndarray) -> np.ndarray:
    """Return the exact PCM format Gradio writes, without its native conversion path."""
    clipped = np.clip(waveform, -1.0, 1.0)
    return np.ascontiguousarray(np.rint(clipped * 32767.0), dtype=np.int16)


def pitch_shift_speech(
    waveform: np.ndarray,
    semitones: float,
    frame_length: int = 1024,
) -> np.ndarray:
    """Shift speech pitch without changing duration using WSOLA and resampling."""
    waveform = np.asarray(waveform, dtype=np.float32)
    if waveform.size < frame_length or abs(semitones) < 0.01:
        return waveform

    factor = 2.0 ** (float(semitones) / 12.0)
    target_stretched_length = max(frame_length, round(len(waveform) * factor))

    # Padding lets WSOLA flush its final overlap without truncating speech.
    reader = ArrayReader(np.pad(waveform, (0, frame_length * 4))[None, :])
    writer = ArrayWriter(1)
    wsola(
        1,
        speed=1.0 / factor,
        frame_length=frame_length,
        tolerance=frame_length // 4,
    ).run(reader, writer)

    stretched = writer.data[0]
    if len(stretched) < target_stretched_length:
        stretched = np.pad(
            stretched,
            (0, target_stretched_length - len(stretched)),
        )
    else:
        stretched = stretched[:target_stretched_length]

    ratio = Fraction(factor).limit_denominator(1000)
    shifted = resample_poly(stretched, ratio.denominator, ratio.numerator)
    if len(shifted) < len(waveform):
        shifted = np.pad(shifted, (0, len(waveform) - len(shifted)))
    return np.asarray(shifted[: len(waveform)], dtype=np.float32)


class Engine:
    def __init__(self, spec: ModelSpec) -> None:
        if not spec.config.is_file():
            raise FileNotFoundError(f"Missing release configuration for {spec.label}")
        if not spec.checkpoint.is_file():
            raise FileNotFoundError(f"Missing release weights for {spec.label}")
        self.spec = spec
        self.hps = utils.get_hparams_from_file(str(spec.config))
        self.model = SynthesizerTrn(
            len(symbols),
            self.hps.data.filter_length // 2 + 1,
            self.hps.train.segment_size // self.hps.data.hop_length,
            **self.hps.model,
        ).eval()
        root_logger = logging.getLogger()
        previous_level = root_logger.level
        try:
            root_logger.setLevel(logging.WARNING)
            utils.load_checkpoint(str(spec.checkpoint), self.model, None)
        finally:
            root_logger.setLevel(previous_level)
        self.device = torch.device("cpu")
        self.sample_rate = int(self.hps.data.sampling_rate)
        self.lock = threading.Lock()

    def tokens(self, text: str) -> tuple[torch.Tensor, torch.Tensor]:
        phonemes = run_vits_frontend(text).phoneme_text
        sequence = cleaned_text_to_sequence(phonemes)
        if self.hps.data.add_blank:
            sequence = commons.intersperse(sequence, 0)
        if not sequence:
            raise ValueError("The phoneme frontend produced no speakable tokens.")
        tokens = torch.LongTensor(sequence).to(self.device).unsqueeze(0)
        lengths = torch.LongTensor([tokens.size(1)]).to(self.device)
        return tokens, lengths

    @torch.inference_mode()
    def synthesize(

        self,

        text: str,

        speed: float,

        variation: float,

        pitch_steps: float,

        seed: int,

    ) -> tuple[int, np.ndarray]:
        chunks = split_text(text)
        waveforms: list[np.ndarray] = []
        with self.lock:
            self.device = torch.device("cuda")
            self.model.to(self.device)
            try:
                for index, chunk in enumerate(chunks):
                    if index:
                        waveforms.append(
                            np.zeros(
                                round(self.sample_rate * boundary_pause_seconds(chunks[index - 1])),
                                dtype=np.float32,
                            )
                        )
                    tokens, lengths = self.tokens(chunk)
                    torch.manual_seed(seed + index)
                    torch.cuda.manual_seed_all(seed + index)
                    waveform = self.model.infer(
                        tokens,
                        lengths,
                        noise_scale=variation,
                        noise_scale_w=0.8,
                        length_scale=1.0 / speed,
                        max_len=4000,
                    )[0][0, 0].float().cpu().numpy()
                    waveforms.append(edge_fade(waveform, self.sample_rate))
            finally:
                self.model.to("cpu")
                self.device = torch.device("cpu")
                torch.cuda.empty_cache()
        audio = np.concatenate(waveforms)
        if abs(pitch_steps) >= 0.01:
            audio = pitch_shift_speech(audio, pitch_steps)
        return self.sample_rate, pcm16(audio)


ENGINES = {label: Engine(spec) for label, spec in SPECS.items()}


def validate(

    text: str,

    speed: float,

    variation: float,

    pitch_steps: float,

    seed: int,

) -> tuple[str, float, float, float, int]:
    text = " ".join((text or "").split())
    if not text:
        raise gr.Error("Enter something for Inflect to say.")
    return text, float(speed), float(variation), float(pitch_steps), int(seed)


@spaces.GPU(duration=120)
def synthesize_one(

    text: str,

    model_name: str,

    speed: float,

    variation: float,

    pitch_steps: float,

    seed: int,

):
    text, speed, variation, pitch_steps, seed = validate(
        text, speed, variation, pitch_steps, seed
    )
    engine = ENGINES[model_name]
    started = time.perf_counter()
    audio = engine.synthesize(text, speed, variation, pitch_steps, seed)
    seconds = len(audio[1]) / audio[0]
    wall = time.perf_counter() - started
    rtf = wall / seconds
    chunks = len(split_text(text))
    return audio, (
        f"**{model_name}** · fixed English voice  \n"
        f"`{engine.spec.params}` parameters · `{seconds:.2f}s` audio · "
        f"`{wall:.2f}s` generation · `RTF {rtf:.3f}` · "
        f"`{len(text)}` characters · `{chunks}` text chunk{'s' if chunks != 1 else ''} · "
        f"pitch `{pitch_steps:+.2f} st` · "
        f"seed `{seed}` · weights `{engine.spec.revision}`"
    )


@spaces.GPU(duration=120)
def synthesize_both(

    text: str,

    speed: float,

    variation: float,

    pitch_steps: float,

    seed: int,

):
    text, speed, variation, pitch_steps, seed = validate(
        text, speed, variation, pitch_steps, seed
    )
    started = time.perf_counter()
    micro = ENGINES["Inflect Micro v2"].synthesize(
        text, speed, variation, pitch_steps, seed
    )
    micro_wall = time.perf_counter() - started
    started = time.perf_counter()
    nano = ENGINES["Inflect Nano v2"].synthesize(
        text, speed, variation, pitch_steps, seed
    )
    nano_wall = time.perf_counter() - started
    micro_seconds = len(micro[1]) / micro[0]
    nano_seconds = len(nano[1]) / nano[0]
    return micro, nano, (
        f"Same text and seed `{seed}` · "
        f"Micro `{micro_wall:.2f}s / {micro_seconds:.2f}s audio` · "
        f"Nano `{nano_wall:.2f}s / {nano_seconds:.2f}s audio`"
    )


CSS = """

:root{color-scheme:dark;--paper:#111412;--panel:#181c19;--panel-2:#141815;--ink:#f3f4ef;--muted:#9ea8a0;--line:#303732;--signal:#2486ff;--signal-hover:#52a0ff;--signal-soft:#15263b}

html,body,.gradio-container,.dark{background:var(--paper)!important;color:var(--ink)!important;color-scheme:dark!important}

body,.gradio-container,.gradio-container button,.gradio-container input,.gradio-container textarea{font-family:Inter,ui-sans-serif,-apple-system,BlinkMacSystemFont,"Segoe UI",sans-serif!important}

.gradio-container{width:min(100%,1640px)!important;max-width:none!important;margin:auto!important;padding:0 clamp(14px,2.2vw,34px) 54px!important;box-sizing:border-box!important}

.gradio-container>.main{padding:0!important}

.gradio-container *{box-sizing:border-box}

.hero-wrap,.hero-wrap.block,.hero-wrap>.wrap{padding:0!important;border:0!important;box-shadow:none!important;background:transparent!important}

.hero{padding:18px 2px 13px;border-bottom:1px solid var(--line);margin-bottom:2px}

.hero-top{display:flex;align-items:center;justify-content:space-between;gap:20px}

.eyebrow{color:var(--signal);font:700 11px/1 ui-monospace,SFMono-Regular,Consolas,monospace;letter-spacing:.13em;text-transform:uppercase}

.links{display:flex;align-items:center;gap:18px;font-size:12px;font-weight:700}

.links a{color:var(--muted)!important;text-decoration:none}.links a:hover{color:var(--ink)!important}

.hero h1{max-width:1040px;margin:15px 0 8px;color:var(--ink);font:650 clamp(40px,4.5vw,56px)/.96 Inter,ui-sans-serif,sans-serif;letter-spacing:-.055em}
.hero>p{max-width:1000px;margin:0;color:var(--muted)!important;font-size:13px;line-height:1.48}
.model-facts{display:flex;align-items:stretch;margin-top:13px;border-top:1px solid var(--line)}
.model-fact{flex:1;padding:8px 24px 0 0;color:var(--muted)!important;font-size:11px;line-height:1.3}
.model-fact+.model-fact{padding-left:20px;border-left:1px solid var(--line)}
.model-fact strong{display:block;margin-bottom:2px;color:var(--ink);font-size:14px;line-height:1.2}
.runtime-tabs{margin:14px 10px 0!important}
.runtime-tabs>.tab-container,.runtime-tabs .tab-container{display:grid!important;grid-template-columns:repeat(2,minmax(0,1fr))!important;gap:0!important;height:auto!important;overflow:hidden!important;border:1px solid var(--line)!important;border-radius:2px!important;background:var(--panel-2)!important}
.runtime-tabs>.tab-container button,.runtime-tabs .tab-container button{display:flex!important;min-height:54px!important;padding:9px 14px!important;flex-direction:column!important;align-items:flex-start!important;justify-content:center!important;border:0!important;border-radius:0!important;background:transparent!important;color:var(--ink)!important;font-size:13px!important;font-weight:800!important;line-height:1.2!important;letter-spacing:.005em!important;transition:background .16s ease,color .16s ease!important}
.runtime-tabs>.tab-container button+button,.runtime-tabs .tab-container button+button{border-left:1px solid var(--line)!important}
.runtime-tabs>.tab-container button::after,.runtime-tabs .tab-container button::after{margin-top:3px;color:var(--muted);font-size:10px;font-weight:550;letter-spacing:.01em}
.runtime-tabs>.tab-container button:nth-child(1)::after,.runtime-tabs .tab-container button:nth-child(1)::after{content:"Hugging Face ZeroGPU"}
.runtime-tabs>.tab-container button:nth-child(2)::after,.runtime-tabs .tab-container button:nth-child(2)::after{content:"WebGPU / WASM · private"}
.runtime-tabs>.tab-container button:hover,.runtime-tabs .tab-container button:hover{background:#1b242d!important;color:var(--ink)!important}
.runtime-tabs>.tab-container button.selected,.runtime-tabs .tab-container button.selected{background:var(--signal-soft)!important;color:#fff!important;box-shadow:inset 0 -2px 0 var(--signal)!important}
.runtime-tabs>.tab-container button.selected::after,.runtime-tabs .tab-container button.selected::after{color:#a9c7e8!important}
.runtime-tabs>.tabitem,.runtime-tabs .tabitem{padding-top:0!important;border:0!important;background:transparent!important}
.mode-banner{display:grid;grid-template-columns:minmax(0,1fr) auto;align-items:center;gap:16px;margin:10px 0;padding:9px 11px;border:1px solid var(--line);border-left:2px solid var(--signal);border-radius:2px;background:var(--panel-2)}
.mode-banner strong{display:block;margin-bottom:2px;color:var(--ink)!important;font-size:12px}
.mode-banner span{color:var(--muted)!important;font-size:11px;line-height:1.45}
.mode-badge{padding:5px 7px;border:1px solid #356ea8;border-radius:2px;color:#dceaff!important;font:750 9px/1 ui-monospace,SFMono-Regular,Consolas,monospace!important;letter-spacing:.08em;text-transform:uppercase;white-space:nowrap}
.browser-frame-shell{margin:0;padding:0;border:1px solid var(--line);border-radius:2px;background:var(--paper);overflow:hidden}
.browser-frame{display:block;width:100%;height:580px;border:0;background:var(--paper);transition:height .18s ease}
.browser-help{margin:9px 2px 0!important;color:var(--muted)!important;font-size:11px!important;line-height:1.5!important}
.workbench{padding:0 0 4px!important;border:0!important;background:transparent!important}
.workbench.block,.workbench>.wrap{border:0!important;box-shadow:none!important;background:transparent!important}
.control-row{align-items:stretch!important;gap:18px!important;margin-top:6px!important;padding:12px 14px 11px!important;border:1px solid var(--line)!important;border-radius:3px!important;background:var(--panel-2)!important}
.control-row>div{min-width:0!important}
.control-row .form{height:100%!important;border:0!important;background:transparent!important}
.control-row label>span:first-child{font-size:12px!important;font-weight:750!important;letter-spacing:.015em!important}
.model-switch .wrap{display:grid!important;grid-template-columns:repeat(2,minmax(0,1fr))!important;gap:0!important;padding:0!important;overflow:hidden!important;border:1px solid var(--line)!important;border-radius:2px!important;background:#101411!important}
.model-switch .wrap label{min-height:46px!important;margin:0!important;padding:10px 12px!important;border:0!important;border-radius:0!important;background:transparent!important}
.model-switch .wrap label+label{border-left:1px solid var(--line)!important}
.model-switch .wrap label:has(input:checked){background:var(--signal-soft)!important;box-shadow:inset 0 -2px 0 var(--signal)!important}
.model-switch .wrap label span{font-size:12px!important;font-weight:750!important}
.action-row{justify-content:center!important;gap:12px!important;margin:12px auto 0!important;max-width:840px!important}
.action-row button{min-height:48px!important}
.comparison{margin-top:22px!important;padding-top:20px!important;border-top:1px solid var(--line)!important}
.comparison-title{margin-bottom:9px!important}
.gradio-container .block,.gradio-container .form{border-radius:3px!important;background:var(--panel-2)!important;border-color:var(--line)!important;color:var(--ink)!important}
.gradio-container .prose,.gradio-container .markdown-body,.gradio-container label,.gradio-container span,.gradio-container p,.gradio-container h1,.gradio-container h2,.gradio-container h3{color:var(--ink)!important}
.gradio-container label{background:transparent!important}
.gradio-container .wrap{background:var(--panel-2)!important}
.gradio-container select{min-height:48px!important;background:#101411!important;color:var(--ink)!important;border-color:var(--line)!important;text-align:left!important}
.gradio-container input,.gradio-container textarea{background:#101411!important;color:var(--ink)!important;border-color:var(--line)!important}
.prompt-input textarea{font-size:16px!important;line-height:1.55!important;min-height:104px!important;padding:15px!important}
.gradio-container button:not(.primary){background:var(--panel-2)!important;color:var(--ink)!important;border-color:var(--line)!important}
.gradio-container button:not(.primary):hover{border-color:#527aa6!important;background:#1b242d!important}
button.primary{min-height:46px!important;border:0!important;border-radius:1px!important;background:var(--signal)!important;color:#fff!important;font-weight:800!important}
button.primary *{color:#fff!important}button.primary:hover{background:var(--signal-hover)!important}
.advanced{margin:10px 0 3px!important;background:#131814!important;border-left:2px solid #344138!important}
.advanced>button{font-size:12px!important;font-weight:750!important;letter-spacing:.01em!important}
.output-audio{margin-top:8px!important}
.outputmeta{padding:11px 13px!important;background:var(--signal-soft)!important;border:1px solid #294e78!important}
.outputmeta *{color:#dceaff!important}
.helper-copy{margin:8px 2px 4px!important;color:var(--muted)!important;font-size:12px!important}
.fineprint{color:var(--muted)!important;font-size:12px;line-height:1.55;border-top:1px solid var(--line);padding-top:16px;margin-top:22px}
footer{display:none!important}
@media(max-width:760px){
html,body{overflow-x:hidden!important}.gradio-container{width:100%!important;min-width:0!important;padding:0 14px 38px!important}
.hero{padding-top:22px}.hero-top{align-items:flex-start;flex-direction:column;gap:12px}.links{gap:14px;font-size:11px}
.hero h1{margin-top:18px;font-size:42px}.model-facts{display:grid;grid-template-columns:1fr 1fr}.model-fact{padding:11px 10px 9px 0}.model-fact+.model-fact{padding-left:10px}.model-fact:nth-child(3){padding-left:0;border-left:0;border-top:1px solid var(--line)}.model-fact:nth-child(4){border-top:1px solid var(--line)}
 .runtime-tabs{margin-left:0!important;margin-right:0!important}.runtime-tabs>.tab-container button,.runtime-tabs .tab-container button{min-height:54px!important;padding-inline:10px!important}.runtime-tabs>.tab-container button::after,.runtime-tabs .tab-container button::after{font-size:9px}.mode-banner{grid-template-columns:1fr}.mode-badge{justify-self:start}.browser-frame{height:980px}
.workbench{min-width:0!important;padding-top:18px!important}.control-row,.action-row,.comparison .row{flex-direction:column!important;min-width:0!important}
.control-row>*,.action-row>*,.comparison .row>*{width:100%!important;min-width:0!important}.gradio-container .wrap,.gradio-container .form,.gradio-container .block{min-width:0!important}
}
"""


PROMPTS = [
    "Wait, are you actually being for real? I thought the train left twenty minutes ago!",
    "Does punctuation work? Yes: commas, semicolons, and periods should create sensible boundaries.",
    "Professor Nguyen mailed the fluorescent poster to Reykjavik on Thursday.",
    "Although the weather changed without warning, the musicians finished packing before anyone opened the enormous glass door.",
]

with gr.Blocks(title="Inflect v2 · local speech playground") as demo:
    gr.HTML("""

    <header class="hero">

      <div class="hero-top">
        <span class="eyebrow">Inflect v2 · choose where inference runs</span>
        <nav class="links">
          <a href="https://huggingface.co/owensong/Inflect-Micro-v2">Micro model</a>
          <a href="https://huggingface.co/owensong/Inflect-Nano-v2">Nano model</a>
          <a href="https://github.com/owenawsong/Inflect">GitHub</a>
        </nav>
      </div>
      <h1>Small models.<br>Complete speech.</h1>
      <p>Generate 24 kHz English speech with the published Inflect-Micro-v2 and Inflect-Nano-v2 checkpoints. Choose hosted ZeroGPU inference or run the ONNX models privately in this browser.</p>
      <div class="model-facts">
        <div class="model-fact"><strong>9.36M · Micro</strong>quality-first model</div>
        <div class="model-fact"><strong>3.96M · Nano</strong>footprint-first model</div>
        <div class="model-fact"><strong>24 kHz WAV</strong>complete waveform output</div>
        <div class="model-fact"><strong>Local stack</strong>text normalization through audio</div>
      </div>
    </header>
    """, elem_classes="hero-wrap")
    with gr.Tabs(elem_classes="runtime-tabs"):
        with gr.Tab("Hosted GPU", id="zerogpu"):
            gr.HTML("""
            <div class="mode-banner">
              <span><strong>ZeroGPU inference</strong>Nothing downloads to your device. Hugging Face runs each request; free-tier quota and queue limits apply.</span>
              <span class="mode-badge">Server mode</span>
            </div>
            """)
            with gr.Column(elem_classes="workbench"):
                text = gr.Textbox(
                    value=PROMPTS[0],
                    lines=4,
                    max_lines=14,
                    label="Text",
                    placeholder="Write a sentence, paragraph, or script...",
                    elem_classes="prompt-input",
                )
                with gr.Row(equal_height=True, elem_classes="control-row"):
                    model = gr.Radio(
                        choices=list(SPECS),
                        value="Inflect Micro v2",
                        label="Model",
                        info="Micro prioritizes quality; Nano prioritizes footprint.",
                        scale=4,
                        elem_classes="model-switch",
                    )
                    speed = gr.Slider(
                        0.7,
                        1.35,
                        value=1.0,
                        step=0.01,
                        label="Speaking speed",
                        info="1.00 is the trained default.",
                        scale=4,
                    )
                    variation = gr.Slider(
                        0.0,
                        1.0,
                        value=0.667,
                        step=0.001,
                        label="Delivery variation",
                        info="Lower is steadier. Higher gives each take more variation.",
                        scale=5,
                    )
                with gr.Accordion(
                    "Advanced controls",
                    open=False,
                    elem_classes="advanced",
                ):
                    with gr.Row():
                        pitch = gr.Slider(
                            -2.0,
                            2.0,
                            value=0.0,
                            step=0.25,
                            label="Pitch shift (semitones)",
                            info="Changes the finished voice pitch without changing speaking speed.",
                        )
                        seed = gr.Number(
                            value=0,
                            precision=0,
                            label="Repeatable seed",
                            info="Use the same seed and settings to reproduce a take.",
                        )
                with gr.Row(elem_classes="action-row"):
                    generate = gr.Button("Generate selected model", variant="primary")
                    compare = gr.Button("Compare Micro and Nano")
                audio = gr.Audio(label="Generated 24 kHz WAV", elem_classes="output-audio")
                status = gr.Markdown("Ready.", elem_classes="outputmeta")
                gr.Markdown(
                    "The first request can include ZeroGPU queue and model-transfer time. "
                    "Use the audio player's menu to download the final WAV.",
                    elem_classes="helper-copy",
                )
                gr.Examples(
                    [[prompt] for prompt in PROMPTS],
                    inputs=text,
                    label="Try a prepared prompt",
                )
            with gr.Column(elem_classes="comparison"):
                gr.Markdown(
                    "### Direct model comparison\n"
                    "The same text, seed, speed, variation, and pitch are sent to both models.",
                    elem_classes="comparison-title",
                )
                with gr.Row():
                    micro_audio = gr.Audio(label="Micro v2 · 9.36M · quality first")
                    nano_audio = gr.Audio(label="Nano v2 · 3.96M · footprint first")
                compare_status = gr.Markdown("", elem_classes="outputmeta")
            gr.HTML("""<p class="fineprint">Fixed English voice · English-only frontend · no voice cloning · no reference audio upload. Uncommon names, abbreviations, and long passages can still fail.</p>""")
        with gr.Tab("This browser", id="webgpu"):
            gr.HTML("""
            <div class="mode-banner">
              <span><strong>Browser-local inference</strong>Your text and generated audio stay in this tab. WebGPU is preferred; compatible browsers fall back to WASM.</span>
              <span class="mode-badge">No queue · no quota</span>
            </div>
            <div class="browser-frame-shell">
              <iframe
                class="browser-frame"
                src="/gradio_api/file=web_runtime/index.html?embed=1"
                title="Inflect browser-local WebGPU runtime"
                allow="autoplay; webgpu"
                loading="eager"
              ></iframe>
            </div>
            <p class="browser-help">The first browser-local run downloads the selected ONNX model and compiles it on your device. Streaming and complete-audio modes are both available inside this panel.</p>
            """, js_on_load="""
            const frame = element.querySelector('.browser-frame');
            if (frame) {
              if (element.__inflectResizeHandler) {
                window.removeEventListener('message', element.__inflectResizeHandler);
              }
              if (element.__inflectLoadHandler) {
                frame.removeEventListener('load', element.__inflectLoadHandler);
              }
              const applyHeight = (height) => {
                const measured = Number(height);
                if (Number.isFinite(measured) && measured > 0) {
                  const responsiveFloor = frame.clientWidth < 1024 ? 980 : 560;
                  frame.style.height = `${Math.max(responsiveFloor, Math.min(2200, measured + 2))}px`;
                }
              };
              const measureFrame = () => {
                try {
                  const doc = frame.contentDocument;
                  applyHeight(Math.max(
                    doc?.body?.scrollHeight || 0,
                    doc?.documentElement?.scrollHeight || 0,
                  ));
                } catch (_) {
                  // The embedded runtime also reports its height by postMessage.
                }
              };
              const resizeHandler = (event) => {
                if (
                  event.origin !== window.location.origin ||
                  event.source !== frame.contentWindow ||
                  event.data?.type !== 'inflect-web-resize'
                ) {
                  return;
                }
                applyHeight(event.data.height);
              };
              element.__inflectResizeHandler = resizeHandler;
              element.__inflectLoadHandler = measureFrame;
              window.addEventListener('message', resizeHandler);
              frame.addEventListener('load', measureFrame);
              [100, 500, 1500].forEach((delay) => window.setTimeout(measureFrame, delay));
            }
            """)
    generate.click(synthesize_one, [text, model, speed, variation, pitch, seed], [audio, status], concurrency_limit=1, api_name="synthesize")
    text.submit(synthesize_one, [text, model, speed, variation, pitch, seed], [audio, status], concurrency_limit=1)
    compare.click(synthesize_both, [text, speed, variation, pitch, seed], [micro_audio, nano_audio, compare_status], concurrency_limit=1, api_name="compare")


if __name__ == "__main__":
    demo.queue(default_concurrency_limit=1).launch(
        theme=gr.themes.Base(),
        css=CSS,
        allowed_paths=[str(WEB_RUNTIME)],
    )