File size: 18,759 Bytes
b2eb230
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import gc
import html
import io
import os
import queue
import wave
from argparse import ArgumentParser
from functools import partial
from pathlib import Path

import gradio as gr
import librosa
import numpy as np
import pyrootutils
import torch
from loguru import logger
from transformers import AutoTokenizer

pyrootutils.setup_root(__file__, indicator=".project-root", pythonpath=True)


from fish_speech.i18n import i18n
from fish_speech.text.chn_text_norm.text import Text as ChnNormedText
from fish_speech.utils import autocast_exclude_mps, set_seed
from tools.api import decode_vq_tokens, encode_reference
from tools.file import AUDIO_EXTENSIONS, list_files
from tools.llama.generate import (
    GenerateRequest,
    GenerateResponse,
    WrappedGenerateResponse,
    launch_thread_safe_queue,
)
from tools.vqgan.inference import load_model as load_decoder_model

# Make einx happy
os.environ["EINX_FILTER_TRACEBACK"] = "false"


HEADER_MD = f"""# Fish Speech



{i18n("A text-to-speech model based on VQ-GAN and Llama developed by [Fish Audio](https://fish.audio).")}  



{i18n("You can find the source code [here](https://github.com/fishaudio/fish-speech) and models [here](https://huggingface.co/fishaudio/fish-speech-1.4).")}  



{i18n("Related code and weights are released under CC BY-NC-SA 4.0 License.")}  



{i18n("We are not responsible for any misuse of the model, please consider your local laws and regulations before using it.")}  

"""

TEXTBOX_PLACEHOLDER = i18n("Put your text here.")
SPACE_IMPORTED = False


def build_html_error_message(error):
    return f"""

    <div style="color: red; 

    font-weight: bold;">

        {html.escape(str(error))}

    </div>

    """


@torch.inference_mode()
def inference(

    text,

    enable_reference_audio,

    reference_audio,

    reference_text,

    max_new_tokens,

    chunk_length,

    top_p,

    repetition_penalty,

    temperature,

    seed="0",

    streaming=False,

):
    if args.max_gradio_length > 0 and len(text) > args.max_gradio_length:
        return (
            None,
            None,
            i18n("Text is too long, please keep it under {} characters.").format(
                args.max_gradio_length
            ),
        )

    seed = int(seed)
    if seed != 0:
        set_seed(seed)
        logger.warning(f"set seed: {seed}")

    # Parse reference audio aka prompt
    prompt_tokens = encode_reference(
        decoder_model=decoder_model,
        reference_audio=reference_audio,
        enable_reference_audio=enable_reference_audio,
    )

    # LLAMA Inference
    request = dict(
        device=decoder_model.device,
        max_new_tokens=max_new_tokens,
        text=text,
        top_p=top_p,
        repetition_penalty=repetition_penalty,
        temperature=temperature,
        compile=args.compile,
        iterative_prompt=chunk_length > 0,
        chunk_length=chunk_length,
        max_length=2048,
        prompt_tokens=prompt_tokens if enable_reference_audio else None,
        prompt_text=reference_text if enable_reference_audio else None,
    )

    response_queue = queue.Queue()
    llama_queue.put(
        GenerateRequest(
            request=request,
            response_queue=response_queue,
        )
    )

    if streaming:
        yield wav_chunk_header(), None, None

    segments = []

    while True:
        result: WrappedGenerateResponse = response_queue.get()
        if result.status == "error":
            yield None, None, build_html_error_message(result.response)
            break

        result: GenerateResponse = result.response
        if result.action == "next":
            break

        with autocast_exclude_mps(
            device_type=decoder_model.device.type, dtype=args.precision
        ):
            fake_audios = decode_vq_tokens(
                decoder_model=decoder_model,
                codes=result.codes,
            )

        fake_audios = fake_audios.float().cpu().numpy()
        segments.append(fake_audios)

        if streaming:
            wav_header = wav_chunk_header()
            audio_data = (fake_audios * 32768).astype(np.int16).tobytes()
            yield wav_header + audio_data, None, None

    if len(segments) == 0:
        return (
            None,
            None,
            build_html_error_message(
                i18n("No audio generated, please check the input text.")
            ),
        )

    # No matter streaming or not, we need to return the final audio
    audio = np.concatenate(segments, axis=0)
    yield None, (decoder_model.spec_transform.sample_rate, audio), None

    if torch.cuda.is_available():
        torch.cuda.empty_cache()
        gc.collect()


inference_stream = partial(inference, streaming=True)

n_audios = 4

global_audio_list = []
global_error_list = []


def inference_wrapper(

    text,

    enable_reference_audio,

    reference_audio,

    reference_text,

    max_new_tokens,

    chunk_length,

    top_p,

    repetition_penalty,

    temperature,

    seed,

    batch_infer_num,

):
    audios = []
    errors = []

    for _ in range(batch_infer_num):
        result = inference(
            text,
            enable_reference_audio,
            reference_audio,
            reference_text,
            max_new_tokens,
            chunk_length,
            top_p,
            repetition_penalty,
            temperature,
            seed,
        )

        _, audio_data, error_message = next(result)

        audios.append(
            gr.Audio(value=audio_data if audio_data else None, visible=True),
        )
        errors.append(
            gr.HTML(value=error_message if error_message else None, visible=True),
        )

    for _ in range(batch_infer_num, n_audios):
        audios.append(
            gr.Audio(value=None, visible=False),
        )
        errors.append(
            gr.HTML(value=None, visible=False),
        )

    return None, *audios, *errors


def wav_chunk_header(sample_rate=44100, bit_depth=16, channels=1):
    buffer = io.BytesIO()

    with wave.open(buffer, "wb") as wav_file:
        wav_file.setnchannels(channels)
        wav_file.setsampwidth(bit_depth // 8)
        wav_file.setframerate(sample_rate)

    wav_header_bytes = buffer.getvalue()
    buffer.close()
    return wav_header_bytes


def normalize_text(user_input, use_normalization):
    if use_normalization:
        return ChnNormedText(raw_text=user_input).normalize()
    else:
        return user_input


def update_examples():
    examples_dir = Path("references")
    examples_dir.mkdir(parents=True, exist_ok=True)
    example_audios = list_files(examples_dir, AUDIO_EXTENSIONS, recursive=True)
    return gr.Dropdown(choices=example_audios + [""])


def build_app():
    with gr.Blocks(theme=gr.themes.Base()) as app:
        gr.Markdown(HEADER_MD)

        # Use light theme by default
        app.load(
            None,
            None,
            js="() => {const params = new URLSearchParams(window.location.search);if (!params.has('__theme')) {params.set('__theme', '%s');window.location.search = params.toString();}}"
            % args.theme,
        )

        # Inference
        with gr.Row():
            with gr.Column(scale=3):
                text = gr.Textbox(
                    label=i18n("Input Text"), placeholder=TEXTBOX_PLACEHOLDER, lines=10
                )
                refined_text = gr.Textbox(
                    label=i18n("Realtime Transform Text"),
                    placeholder=i18n(
                        "Normalization Result Preview (Currently Only Chinese)"
                    ),
                    lines=5,
                    interactive=False,
                )

                with gr.Row():
                    if_refine_text = gr.Checkbox(
                        label=i18n("Text Normalization"),
                        value=False,
                        scale=1,
                    )

                with gr.Row():
                    with gr.Column():
                        with gr.Tab(label=i18n("Advanced Config")):
                            with gr.Row():
                                chunk_length = gr.Slider(
                                    label=i18n("Iterative Prompt Length, 0 means off"),
                                    minimum=50,
                                    maximum=300,
                                    value=200,
                                    step=8,
                                )

                                max_new_tokens = gr.Slider(
                                    label=i18n(
                                        "Maximum tokens per batch, 0 means no limit"
                                    ),
                                    minimum=0,
                                    maximum=2048,
                                    value=0,  # 0 means no limit
                                    step=8,
                                )

                            with gr.Row():
                                top_p = gr.Slider(
                                    label="Top-P",
                                    minimum=0.6,
                                    maximum=0.9,
                                    value=0.7,
                                    step=0.01,
                                )

                                repetition_penalty = gr.Slider(
                                    label=i18n("Repetition Penalty"),
                                    minimum=1,
                                    maximum=1.5,
                                    value=1.2,
                                    step=0.01,
                                )

                            with gr.Row():
                                temperature = gr.Slider(
                                    label="Temperature",
                                    minimum=0.6,
                                    maximum=0.9,
                                    value=0.7,
                                    step=0.01,
                                )
                                seed = gr.Textbox(
                                    label="Seed",
                                    info="0 means randomized inference, otherwise deterministic",
                                    placeholder="any 32-bit-integer",
                                    value="0",
                                )

                        with gr.Tab(label=i18n("Reference Audio")):
                            with gr.Row():
                                gr.Markdown(
                                    i18n(
                                        "5 to 10 seconds of reference audio, useful for specifying speaker."
                                    )
                                )
                            with gr.Row():
                                enable_reference_audio = gr.Checkbox(
                                    label=i18n("Enable Reference Audio"),
                                )

                            with gr.Row():
                                example_audio_dropdown = gr.Dropdown(
                                    label=i18n("Select Example Audio"),
                                    choices=[""],
                                    value="",
                                    interactive=True,
                                    allow_custom_value=True,
                                )
                            with gr.Row():
                                reference_audio = gr.Audio(
                                    label=i18n("Reference Audio"),
                                    type="filepath",
                                )
                            with gr.Row():
                                reference_text = gr.Textbox(
                                    label=i18n("Reference Text"),
                                    lines=1,
                                    placeholder="在一无所知中,梦里的一天结束了,一个新的「轮回」便会开始。",
                                    value="",
                                )
                        with gr.Tab(label=i18n("Batch Inference")):
                            with gr.Row():
                                batch_infer_num = gr.Slider(
                                    label="Batch infer nums",
                                    minimum=1,
                                    maximum=n_audios,
                                    step=1,
                                    value=1,
                                )

            with gr.Column(scale=3):
                for _ in range(n_audios):
                    with gr.Row():
                        error = gr.HTML(
                            label=i18n("Error Message"),
                            visible=True if _ == 0 else False,
                        )
                        global_error_list.append(error)
                    with gr.Row():
                        audio = gr.Audio(
                            label=i18n("Generated Audio"),
                            type="numpy",
                            interactive=False,
                            visible=True if _ == 0 else False,
                        )
                        global_audio_list.append(audio)

                with gr.Row():
                    stream_audio = gr.Audio(
                        label=i18n("Streaming Audio"),
                        streaming=True,
                        autoplay=True,
                        interactive=False,
                        show_download_button=True,
                    )
                with gr.Row():
                    with gr.Column(scale=3):
                        generate = gr.Button(
                            value="\U0001F3A7 " + i18n("Generate"), variant="primary"
                        )
                        generate_stream = gr.Button(
                            value="\U0001F3A7 " + i18n("Streaming Generate"),
                            variant="primary",
                        )

        text.input(
            fn=normalize_text, inputs=[text, if_refine_text], outputs=[refined_text]
        )

        def select_example_audio(audio_path):
            audio_path = Path(audio_path)
            if audio_path.is_file():
                lab_file = Path(audio_path.with_suffix(".lab"))

                if lab_file.exists():
                    lab_content = lab_file.read_text(encoding="utf-8").strip()
                else:
                    lab_content = ""

                return str(audio_path), lab_content, True
            return None, "", False

        # Connect the dropdown to update reference audio and text
        example_audio_dropdown.change(
            fn=update_examples, inputs=[], outputs=[example_audio_dropdown]
        ).then(
            fn=select_example_audio,
            inputs=[example_audio_dropdown],
            outputs=[reference_audio, reference_text, enable_reference_audio],
        )

        # # Submit
        generate.click(
            inference_wrapper,
            [
                refined_text,
                enable_reference_audio,
                reference_audio,
                reference_text,
                max_new_tokens,
                chunk_length,
                top_p,
                repetition_penalty,
                temperature,
                seed,
                batch_infer_num,
            ],
            [stream_audio, *global_audio_list, *global_error_list],
            concurrency_limit=1,
        )

        generate_stream.click(
            inference_stream,
            [
                refined_text,
                enable_reference_audio,
                reference_audio,
                reference_text,
                max_new_tokens,
                chunk_length,
                top_p,
                repetition_penalty,
                temperature,
                seed,
            ],
            [stream_audio, global_audio_list[0], global_error_list[0]],
            concurrency_limit=1,
        )
    return app


def parse_args():
    parser = ArgumentParser()
    parser.add_argument(
        "--llama-checkpoint-path",
        type=Path,
        default="checkpoints/fish-speech-1.4",
    )
    parser.add_argument(
        "--decoder-checkpoint-path",
        type=Path,
        default="checkpoints/fish-speech-1.4/firefly-gan-vq-fsq-8x1024-21hz-generator.pth",
    )
    parser.add_argument("--decoder-config-name", type=str, default="firefly_gan_vq")
    parser.add_argument("--device", type=str, default="cuda")
    parser.add_argument("--half", action="store_true")
    parser.add_argument("--compile", action="store_true")
    parser.add_argument("--max-gradio-length", type=int, default=0)
    parser.add_argument("--theme", type=str, default="light")

    return parser.parse_args()


if __name__ == "__main__":
    args = parse_args()
    args.precision = torch.half if args.half else torch.bfloat16

    # Check if CUDA is available
    if not torch.cuda.is_available():
        logger.info("CUDA is not available, running on CPU.")
        args.device = "cpu"

    logger.info("Loading Llama model...")
    llama_queue = launch_thread_safe_queue(
        checkpoint_path=args.llama_checkpoint_path,
        device=args.device,
        precision=args.precision,
        compile=args.compile,
    )
    logger.info("Llama model loaded, loading VQ-GAN model...")

    decoder_model = load_decoder_model(
        config_name=args.decoder_config_name,
        checkpoint_path=args.decoder_checkpoint_path,
        device=args.device,
    )

    logger.info("Decoder model loaded, warming up...")

    # Dry run to check if the model is loaded correctly and avoid the first-time latency
    list(
        inference(
            text="Hello, world!",
            enable_reference_audio=False,
            reference_audio=None,
            reference_text="",
            max_new_tokens=0,
            chunk_length=200,
            top_p=0.7,
            repetition_penalty=1.2,
            temperature=0.7,
        )
    )

    logger.info("Warming up done, launching the web UI...")

    app = build_app()
    app.launch(show_api=True)