|
import gradio as gr |
|
from openai import OpenAI |
|
import requests |
|
import json |
|
|
|
import torch |
|
import torchaudio |
|
from einops import rearrange |
|
import argparse |
|
import json |
|
import os |
|
import spaces |
|
from tqdm import tqdm |
|
import random |
|
import numpy as np |
|
import sys |
|
import base64 |
|
from diffrhythm.infer.infer_utils import ( |
|
get_reference_latent, |
|
get_lrc_token, |
|
get_audio_style_prompt, |
|
get_text_style_prompt, |
|
prepare_model, |
|
get_negative_style_prompt |
|
) |
|
from diffrhythm.infer.infer import inference |
|
|
|
MAX_SEED = np.iinfo(np.int32).max |
|
device='cuda' |
|
cfm, cfm_full, tokenizer, muq, vae = prepare_model(device) |
|
cfm = torch.compile(cfm) |
|
cfm_full = torch.compile(cfm_full) |
|
|
|
@spaces.GPU(duration=40) |
|
def infer_music(lrc, ref_audio_path, text_prompt, current_prompt_type, seed=42, randomize_seed=False, steps=32, cfg_strength=4.0, file_type='wav', odeint_method='euler', Music_Duration='95s', device='cuda'): |
|
if Music_Duration == '95s': |
|
max_frames = 2048 |
|
cfm_model = cfm |
|
else: |
|
max_frames = 6144 |
|
cfm_model = cfm_full |
|
if randomize_seed: |
|
seed = random.randint(0, MAX_SEED) |
|
torch.manual_seed(seed) |
|
sway_sampling_coef = -1 if steps < 32 else None |
|
vocal_flag = False |
|
try: |
|
lrc_prompt, start_time = get_lrc_token(max_frames, lrc, tokenizer, device) |
|
if current_prompt_type == 'audio': |
|
style_prompt, vocal_flag = get_audio_style_prompt(muq, ref_audio_path) |
|
else: |
|
style_prompt = get_text_style_prompt(muq, text_prompt) |
|
except Exception as e: |
|
raise gr.Error(f"Error: {str(e)}") |
|
negative_style_prompt = get_negative_style_prompt(device) |
|
latent_prompt = get_reference_latent(device, max_frames) |
|
generated_song = inference(cfm_model=cfm_model, |
|
vae_model=vae, |
|
cond=latent_prompt, |
|
text=lrc_prompt, |
|
duration=max_frames, |
|
style_prompt=style_prompt, |
|
negative_style_prompt=negative_style_prompt, |
|
steps=steps, |
|
cfg_strength=cfg_strength, |
|
sway_sampling_coef=sway_sampling_coef, |
|
start_time=start_time, |
|
file_type=file_type, |
|
vocal_flag=vocal_flag, |
|
odeint_method=odeint_method, |
|
) |
|
return generated_song |
|
|
|
def R1_infer1(theme, tags_gen, language): |
|
try: |
|
client = OpenAI(api_key=os.getenv('HS_DP_API'), base_url = "https://ark.cn-beijing.volces.com/api/v3") |
|
|
|
llm_prompt = """ |
|
请围绕"{theme}"主题生成一首符合"{tags}"风格的语言为{language}的完整歌词。严格遵循以下要求： |
|
|
|
### **强制格式规则** |
|
1. **仅输出时间戳和歌词**，禁止任何括号、旁白、段落标记（如副歌、间奏、尾奏等注释）。 |
|
2. 每行格式必须为 `[mm:ss.xx]歌词内容`，时间戳与歌词间无空格，歌词内容需完整连贯。 |
|
3. 时间戳需自然分布，**第一句歌词起始时间不得为 [00:00.00]**，需考虑前奏空白。 |
|
|
|
### **内容与结构要求** |
|
1. 歌词应富有变化，使情绪递进，整体连贯有层次感。**每行歌词长度应自然变化**，切勿长度一致，导致很格式化。 |
|
2. **时间戳分配应根据歌曲的标签、歌词的情感、节奏来合理推测**，而非机械地按照歌词长度分配。 |
|
3. 间奏/尾奏仅通过时间空白体现（如从 [02:30.00] 直接跳至 [02:50.00]），**无需文字描述**。 |
|
|
|
### **负面示例（禁止出现）** |
|
- 错误：[01:30.00](钢琴间奏) |
|
- 错误：[02:00.00][副歌] |
|
- 错误：空行、换行符、注释 |
|
""" |
|
|
|
response = client.chat.completions.create( |
|
model="ep-20250304144033-nr9wl", |
|
messages=[ |
|
{"role": "system", "content": "You are a professional musician who has been invited to make music-related comments."}, |
|
{"role": "user", "content": llm_prompt.format(theme=theme, tags=tags_gen, language=language)}, |
|
], |
|
stream=False |
|
) |
|
|
|
info = response.choices[0].message.content |
|
|
|
return info |
|
|
|
except requests.exceptions.RequestException as e: |
|
print(f'请求出错: {e}') |
|
return {} |
|
|
|
|
|
|
|
def R1_infer2(tags_lyrics, lyrics_input): |
|
client = OpenAI(api_key=os.getenv('HS_DP_API'), base_url = "https://ark.cn-beijing.volces.com/api/v3") |
|
|
|
llm_prompt = """ |
|
{lyrics_input}这是一首歌的歌词,每一行是一句歌词,{tags_lyrics}是我希望这首歌的风格，我现在想要给这首歌的每一句歌词打时间戳得到LRC，我希望时间戳分配应根据歌曲的标签、歌词的情感、节奏来合理推测，而非机械地按照歌词长度分配。第一句歌词的时间戳应考虑前奏长度，避免歌词从 `[00:00.00]` 直接开始。严格按照 LRC 格式输出歌词，每行格式为 `[mm:ss.xx]歌词内容`。最后的结果只输出LRC,不需要其他的解释。 |
|
""" |
|
|
|
response = client.chat.completions.create( |
|
model="ep-20250304144033-nr9wl", |
|
messages=[ |
|
{"role": "system", "content": "You are a professional musician who has been invited to make music-related comments."}, |
|
{"role": "user", "content": llm_prompt.format(lyrics_input=lyrics_input, tags_lyrics=tags_lyrics)}, |
|
], |
|
stream=False |
|
) |
|
|
|
info = response.choices[0].message.content |
|
|
|
return info |
|
|
|
css = """ |
|
/* 固定文本域高度并强制滚动条 */ |
|
.lyrics-scroll-box textarea { |
|
height: 405px !important; /* 固定高度 */ |
|
max-height: 500px !important; /* 最大高度 */ |
|
overflow-y: auto !important; /* 垂直滚动 */ |
|
white-space: pre-wrap; /* 保留换行 */ |
|
line-height: 1.5; /* 行高优化 */ |
|
} |
|
|
|
.gr-examples { |
|
background: transparent !important; |
|
border: 1px solid #e0e0e0 !important; |
|
border-radius: 8px; |
|
margin: 1rem 0 !important; |
|
padding: 1rem !important; |
|
} |
|
|
|
""" |
|
|
|
|
|
with gr.Blocks(css=css) as demo: |
|
gr.HTML(f""" |
|
<div style="display: flex; align-items: center;"> |
|
<img src='https://huggingface.co/spaces/demohug/demo11213/resolve/main/melodyVerseLogo.svg' |
|
style='width: 200px; height: 40%; display: block; margin: 0 auto 20px;'> |
|
</div> |
|
|
|
<div style="flex: 1; text-align: center;"> |
|
<div style="font-size: 2em; font-weight: bold; text-align: center; margin-bottom: 5px"> |
|
MelodyVerse |
|
</div> |
|
</div> |
|
""") |
|
|
|
with gr.Tabs() as tabs: |
|
|
|
|
|
with gr.Tab("Music Generate", id=0): |
|
with gr.Row(): |
|
with gr.Column(): |
|
lrc = gr.Textbox( |
|
label="Lyrics", |
|
placeholder="Input the full lyrics", |
|
lines=12, |
|
max_lines=50, |
|
elem_classes="lyrics-scroll-box", |
|
value="""[00:04.34]Tell me that I'm special\n[00:06.57]Tell me I look pretty\n[00:08.46]Tell me I'm a little angel\n[00:10.58]Sweetheart of your city\n[00:13.64]Say what I'm dying to hear\n[00:17.35]Cause I'm dying to hear you\n[00:20.86]Tell me I'm that new thing\n[00:22.93]Tell me that I'm relevant\n[00:24.96]Tell me that I got a big heart\n[00:27.04]Then back it up with evidence\n[00:29.94]I need it and I don't know why\n[00:34.28]This late at night\n[00:36.32]Isn't it lonely\n[00:39.24]I'd do anything to make you want me\n[00:43.40]I'd give it all up if you told me\n[00:47.42]That I'd be\n[00:49.43]The number one girl in your eyes\n[00:52.85]Your one and only\n[00:55.74]So what's it gon' take for you to want me\n[00:59.78]I'd give it all up if you told me\n[01:03.89]That I'd be\n[01:05.94]The number one girl in your eyes\n[01:11.34]Tell me I'm going real big places\n[01:14.32]Down to earth so friendly\n[01:16.30]And even through all the phases\n[01:18.46]Tell me you accept me\n[01:21.56]Well that's all I'm dying to hear\n[01:25.30]Yeah I'm dying to hear you\n[01:28.91]Tell me that you need me\n[01:30.85]Tell me that I'm loved\n[01:32.90]Tell me that I'm worth it\n[01:34.95]And that I'm enough\n[01:37.91]I need it and I don't know why\n[01:42.08]This late at night\n[01:44.24]Isn't it lonely\n[01:47.18]I'd do anything to make you want me\n[01:51.30]I'd give it all up if you told me\n[01:55.32]That I'd be\n[01:57.35]The number one girl in your eyes\n[02:00.72]Your one and only\n[02:03.57]So what's it gon' take for you to want me\n[02:07.78]I'd give it all up if you told me\n[02:11.74]That I'd be\n[02:13.86]The number one girl in your eyes\n[02:17.03]The girl in your eyes\n[02:21.05]The girl in your eyes\n[02:26.30]Tell me I'm the number one girl\n[02:28.44]I'm the number one girl in your eyes\n[02:33.49]The girl in your eyes\n[02:37.58]The girl in your eyes\n[02:42.74]Tell me I'm the number one girl\n[02:44.88]I'm the number one girl in your eyes\n[02:49.91]Well isn't it lonely\n[02:53.19]I'd do anything to make you want me\n[02:57.10]I'd give it all up if you told me\n[03:01.15]That I'd be\n[03:03.31]The number one girl in your eyes\n[03:06.57]Your one and only\n[03:09.42]So what's it gon' take for you to want me\n[03:13.50]I'd give it all up if you told me\n[03:17.56]That I'd be\n[03:19.66]The number one girl in your eyes\n[03:25.74]The number one girl in your eyes""" |
|
) |
|
|
|
current_prompt_type = gr.State(value="audio") |
|
with gr.Tabs() as inside_tabs: |
|
with gr.Tab("Audio Prompt"): |
|
audio_prompt = gr.Audio(label="Audio Prompt", type="filepath", value="./src/prompt/default.wav") |
|
with gr.Tab("Text Prompt"): |
|
text_prompt = gr.Textbox( |
|
label="Text Prompt", |
|
placeholder="Enter the Text Prompt, eg: emotional piano pop", |
|
) |
|
def update_prompt_type(evt: gr.SelectData): |
|
return "audio" if evt.index == 0 else "text" |
|
|
|
inside_tabs.select( |
|
fn=update_prompt_type, |
|
outputs=current_prompt_type |
|
) |
|
|
|
with gr.Column(): |
|
with gr.Accordion("Best Practices Guide", open=True): |
|
gr.Markdown(""" |
|
1. **Lyrics Format Requirements** |
|
- Each line must follow: `[mm:ss.xx]Lyric content` |
|
- Example of valid format: |
|
``` |
|
[00:10.00]Moonlight spills through broken blinds |
|
[00:13.20]Your shadow dances on the dashboard shrine |
|
``` |
|
|
|
2. **Audio Prompt Requirements** |
|
- Reference audio should be ≥ 1 second, audio >10 seconds will be randomly clipped into 10 seconds |
|
- For optimal results, the 10-second clips should be carefully selected |
|
- Shorter clips may lead to incoherent generation |
|
3. **Supported Languages** |
|
- **Chinese and English** |
|
- More languages comming soon |
|
|
|
4. **Others** |
|
- If loading audio result is slow, you can select Output Format as mp3 in Advanced Settings. |
|
|
|
""") |
|
Music_Duration = gr.Radio(["95s", "285s"], label="Music Duration", value="95s") |
|
|
|
lyrics_btn = gr.Button("Generate", variant="primary") |
|
audio_output = gr.Audio(label="Audio Result", type="filepath", elem_id="audio_output") |
|
with gr.Accordion("Advanced Settings", open=False): |
|
seed = gr.Slider( |
|
label="Seed", |
|
minimum=0, |
|
maximum=MAX_SEED, |
|
step=1, |
|
value=0, |
|
) |
|
randomize_seed = gr.Checkbox(label="Randomize seed", value=True) |
|
|
|
steps = gr.Slider( |
|
minimum=10, |
|
maximum=100, |
|
value=32, |
|
step=1, |
|
label="Diffusion Steps", |
|
interactive=True, |
|
elem_id="step_slider" |
|
) |
|
cfg_strength = gr.Slider( |
|
minimum=1, |
|
maximum=10, |
|
value=4.0, |
|
step=0.5, |
|
label="CFG Strength", |
|
interactive=True, |
|
elem_id="step_slider" |
|
) |
|
odeint_method = gr.Radio(["euler", "midpoint", "rk4","implicit_adams"], label="ODE Solver", value="euler") |
|
file_type = gr.Dropdown(["wav", "mp3", "ogg"], label="Output Format", value="wav") |
|
|
|
|
|
gr.Examples( |
|
examples=[ |
|
["./src/prompt/pop_cn.wav"], |
|
["./src/prompt/pop_en.wav"], |
|
["./src/prompt/rock_cn.wav"], |
|
["./src/prompt/rock_en.wav"], |
|
["./src/prompt/country_cn.wav"], |
|
["./src/prompt/country_en.wav"], |
|
["./src/prompt/classic_cn.wav"], |
|
["./src/prompt/classic_en.wav"], |
|
["./src/prompt/jazz_cn.wav"], |
|
["./src/prompt/jazz_en.wav"], |
|
["./src/prompt/rap_cn.wav"], |
|
["./src/prompt/rap_en.wav"], |
|
["./src/prompt/default.wav"] |
|
], |
|
inputs=[audio_prompt], |
|
label="Audio Examples", |
|
examples_per_page=13, |
|
elem_id="audio-examples-container" |
|
) |
|
|
|
gr.Examples( |
|
examples=[ |
|
["Pop Emotional Piano"], |
|
["流行 情感 钢琴"], |
|
["Indie folk ballad, coming-of-age themes, acoustic guitar picking with harmonica interludes"], |
|
["独立民谣, 成长主题, 原声吉他弹奏与口琴间奏"] |
|
], |
|
inputs=[text_prompt], |
|
label="Text Examples", |
|
examples_per_page=4, |
|
elem_id="text-examples-container" |
|
) |
|
|
|
gr.Examples( |
|
examples=[ |
|
["""[00:04.34]Tell me that I'm special\n[00:06.57]Tell me I look pretty\n[00:08.46]Tell me I'm a little angel\n[00:10.58]Sweetheart of your city\n[00:13.64]Say what I'm dying to hear\n[00:17.35]Cause I'm dying to hear you\n[00:20.86]Tell me I'm that new thing\n[00:22.93]Tell me that I'm relevant\n[00:24.96]Tell me that I got a big heart\n[00:27.04]Then back it up with evidence\n[00:29.94]I need it and I don't know why\n[00:34.28]This late at night\n[00:36.32]Isn't it lonely\n[00:39.24]I'd do anything to make you want me\n[00:43.40]I'd give it all up if you told me\n[00:47.42]That I'd be\n[00:49.43]The number one girl in your eyes\n[00:52.85]Your one and only\n[00:55.74]So what's it gon' take for you to want me\n[00:59.78]I'd give it all up if you told me\n[01:03.89]That I'd be\n[01:05.94]The number one girl in your eyes\n[01:11.34]Tell me I'm going real big places\n[01:14.32]Down to earth so friendly\n[01:16.30]And even through all the phases\n[01:18.46]Tell me you accept me\n[01:21.56]Well that's all I'm dying to hear\n[01:25.30]Yeah I'm dying to hear you\n[01:28.91]Tell me that you need me\n[01:30.85]Tell me that I'm loved\n[01:32.90]Tell me that I'm worth it\n[01:34.95]And that I'm enough\n[01:37.91]I need it and I don't know why\n[01:42.08]This late at night\n[01:44.24]Isn't it lonely\n[01:47.18]I'd do anything to make you want me\n[01:51.30]I'd give it all up if you told me\n[01:55.32]That I'd be\n[01:57.35]The number one girl in your eyes\n[02:00.72]Your one and only\n[02:03.57]So what's it gon' take for you to want me\n[02:07.78]I'd give it all up if you told me\n[02:11.74]That I'd be\n[02:13.86]The number one girl in your eyes\n[02:17.03]The girl in your eyes\n[02:21.05]The girl in your eyes\n[02:26.30]Tell me I'm the number one girl\n[02:28.44]I'm the number one girl in your eyes\n[02:33.49]The girl in your eyes\n[02:37.58]The girl in your eyes\n[02:42.74]Tell me I'm the number one girl\n[02:44.88]I'm the number one girl in your eyes\n[02:49.91]Well isn't it lonely\n[02:53.19]I'd do anything to make you want me\n[02:57.10]I'd give it all up if you told me\n[03:01.15]That I'd be\n[03:03.31]The number one girl in your eyes\n[03:06.57]Your one and only\n[03:09.42]So what's it gon' take for you to want me\n[03:13.50]I'd give it all up if you told me\n[03:17.56]That I'd be\n[03:19.66]The number one girl in your eyes\n[03:25.74]The number one girl in your eyes"""], |
|
["""[00:00.52]Abracadabra abracadabra\n[00:03.97]Ha\n[00:04.66]Abracadabra abracadabra\n[00:12.02]Yeah\n[00:15.80]Pay the toll to the angels\n[00:19.08]Drawin' circles in the clouds\n[00:23.31]Keep your mind on the distance\n[00:26.67]When the devil turns around\n[00:30.95]Hold me in your heart tonight\n[00:34.11]In the magic of the dark moonlight\n[00:38.44]Save me from this empty fight\n[00:43.83]In the game of life\n[00:45.84]Like a poem said by a lady in red\n[00:49.45]You hear the last few words of your life\n[00:53.15]With a haunting dance now you're both in a trance\n[00:56.90]It's time to cast your spell on the night\n[01:01.40]Abracadabra ama-ooh-na-na\n[01:04.88]Abracadabra porta-ooh-ga-ga\n[01:08.92]Abracadabra abra-ooh-na-na\n[01:12.30]In her tongue she's sayin'\n[01:14.76]Death or love tonight\n[01:18.61]Abracadabra abracadabra\n[01:22.18]Abracadabra abracadabra\n[01:26.08]Feel the beat under your feet\n[01:27.82]The floor's on fire\n[01:29.90]Abracadabra abracadabra\n[01:33.78]Choose the road on the west side\n[01:37.09]As the dust flies watch it burn\n[01:41.45]Don't waste time on feeling\n[01:44.64]Your depression won't return\n[01:49.15]Hold me in your heart tonight\n[01:52.21]In the magic of the dark moonlight\n[01:56.54]Save me from this empty fight\n[02:01.77]In the game of life\n[02:03.94]Like a poem said by a lady in red\n[02:07.52]You hear the last few words of your life\n[02:11.19]With a haunting dance now you're both in a trance\n[02:14.95]It's time to cast your spell on the night\n[02:19.53]Abracadabra ama-ooh-na-na\n[02:22.71]Abracadabra porta-ooh-ga-ga\n[02:26.94]Abracadabra abra-ooh-na-na\n[02:30.42]In her tongue she's sayin'\n[02:32.83]Death or love tonight\n[02:36.55]Abracadabra abracadabra\n[02:40.27]Abracadabra abracadabra\n[02:44.19]Feel the beat under your feet\n[02:46.14]The floor's on fire\n[02:47.95]Abracadabra abracadabra\n[02:51.17]Phantom of the dance floor come to me\n[02:58.46]Sing for me a sinful melody\n[03:06.51]Ah-ah-ah-ah-ah ah-ah ah-ah\n[03:13.76]Ah-ah-ah-ah-ah ah-ah ah-ah\n[03:22.39]Abracadabra ama-ooh-na-na\n[03:25.66]Abracadabra porta-ooh-ga-ga\n[03:29.87]Abracadabra abra-ooh-na-na\n[03:33.16]In her tongue she's sayin'\n[03:35.55]Death or love tonight"""] |
|
], |
|
|
|
inputs=[lrc], |
|
label="Lrc Examples", |
|
examples_per_page=2, |
|
elem_id="lrc-examples-container", |
|
) |
|
|
|
|
|
|
|
with gr.Tab("Lyrics Generate", id=1): |
|
with gr.Row(): |
|
with gr.Column(): |
|
with gr.Accordion("Notice", open=False): |
|
gr.Markdown("**Two Generation Modes:**\n1. Generate from theme & tags\n2. Add timestamps to existing lyrics") |
|
|
|
with gr.Group(): |
|
gr.Markdown("### Method 1: Generate from Theme") |
|
theme = gr.Textbox(label="theme", placeholder="Enter song theme, e.g: Love and Heartbreak") |
|
tags_gen = gr.Textbox(label="tags", placeholder="Enter song tags, e.g: pop confidence healing") |
|
language = gr.Radio(["cn", "en"], label="Language", value="en") |
|
gen_from_theme_btn = gr.Button("Generate LRC (From Theme)", variant="primary") |
|
|
|
gr.Examples( |
|
examples=[ |
|
[ |
|
"Love and Heartbreak", |
|
"vocal emotional piano pop", |
|
"en" |
|
], |
|
[ |
|
"Heroic Epic", |
|
"choir orchestral powerful", |
|
"cn" |
|
] |
|
], |
|
inputs=[theme, tags_gen, language], |
|
label="Examples: Generate from Theme" |
|
) |
|
|
|
with gr.Group(visible=True): |
|
gr.Markdown("### Method 2: Add Timestamps to Lyrics") |
|
tags_lyrics = gr.Textbox(label="tags", placeholder="Enter song tags, e.g: ballad piano slow") |
|
lyrics_input = gr.Textbox( |
|
label="Raw Lyrics (without timestamps)", |
|
placeholder="Enter plain lyrics (without timestamps), e.g:\nYesterday\nAll my troubles...", |
|
lines=10, |
|
max_lines=50, |
|
elem_classes="lyrics-scroll-box" |
|
) |
|
|
|
gen_from_lyrics_btn = gr.Button("Generate LRC (From Lyrics)", variant="primary") |
|
|
|
gr.Examples( |
|
examples=[ |
|
[ |
|
"acoustic folk happy", |
|
"""I'm sitting here in the boring room\nIt's just another rainy Sunday afternoon""" |
|
], |
|
[ |
|
"electronic dance energetic", |
|
"""We're living in a material world\nAnd I am a material girl""" |
|
] |
|
], |
|
inputs=[tags_lyrics, lyrics_input], |
|
label="Examples: Generate from Lyrics" |
|
) |
|
|
|
|
|
with gr.Column(): |
|
lrc_output = gr.Textbox( |
|
label="Generated LRC", |
|
placeholder="Timed lyrics will appear here", |
|
lines=57, |
|
elem_classes="lrc-output", |
|
show_copy_button=True |
|
) |
|
|
|
|
|
gen_from_theme_btn.click( |
|
fn=R1_infer1, |
|
inputs=[theme, tags_gen, language], |
|
outputs=lrc_output |
|
) |
|
|
|
gen_from_lyrics_btn.click( |
|
fn=R1_infer2, |
|
inputs=[tags_lyrics, lyrics_input], |
|
outputs=lrc_output |
|
) |
|
|
|
tabs.select( |
|
lambda s: None, |
|
None, |
|
None |
|
) |
|
|
|
lyrics_btn.click( |
|
fn=infer_music, |
|
inputs=[lrc, audio_prompt, text_prompt, current_prompt_type, seed, randomize_seed, steps, cfg_strength, file_type, odeint_method, Music_Duration], |
|
outputs=audio_output |
|
) |
|
|
|
|
|
if __name__ == "__main__": |
|
demo.launch() |
