Parler-TTS \u9879\u76ee\u5f88\u9ad8\u5174\u5730\u5ba3\u5e03\u53d1\u5e03\u4e24\u4e2a\u65b0\u7684\u6587\u672c\u8f6c\u8bed\u97f3\u6a21\u578b\uff01\u672c\u6587\u4ecb\u7ecd\u5982\u4f55\u4f7f\u7528Gemma\u5236\u4f5c\u81ea\u5df1\u7684\u6570\u636e\u96c6\uff0c\u5982\u4f55\u5fae\u8c03Parler-TTS\uff0c\u4ee5\u53ca\u5982\u4f55\u4f7f\u7528Parler-TTS\u6a21\u578b\u8fdb\u884c\u63a8\u7406\u3002<\/p>\n
\u9996\u5148\uff0c\u6211\u4eec\u6709 \uff0c\u8fd9\u662f\u4e00\u6b3e\u8f7b\u91cf\u7ea7\u6a21\u578b\uff0c\u975e\u5e38\u9002\u5408\u5feb\u901f\u8f7b\u677e\u5730\u751f\u6210\u8bed\u97f3\u3002\u53d7\u6700\u8fd1\u7814\u7a76\u8bba\u6587\u300a\u4f7f\u7528\u5408\u6210\u6807\u6ce8\u7684\u9ad8\u4fdd\u771f\u6587\u672c\u8f6c\u8bed\u97f3\u7684\u81ea\u7136\u8bed\u8a00\u6307\u5bfc\u300b\u7684\u542f\u53d1\uff0cParler-TTS Mini v0.1 \u8ba9\u4f60\u901a\u8fc7\u7b80\u5355\u7684\u6587\u672c\u63d0\u793a\u76f4\u89c2\u5730\u63a7\u5236\u5404\u79cd\u8bed\u97f3\u65b9\u9762\uff0c\u4f8b\u5982\u6027\u522b\u3001\u80cc\u666f\u566a\u97f3\u548c\u8bed\u901f\u3002<\/p>\n
\u5bf9\u4e8e\u90a3\u4e9b\u5bfb\u6c42\u6700\u5927\u8868\u73b0\u529b\u548c\u63a7\u5236\u529b\u7684\u4eba\uff0c\u6211\u4eec\u8fd8\u6709 \u3002\u8fd9\u4e2a 2.2B \u53c2\u6570\u6a21\u578b\u7ecf\u8fc7\u5927\u91cf 45K \u5c0f\u65f6\u7684\u97f3\u9891\u6570\u636e\u8bad\u7ec3\uff0c\u53ef\u63d0\u4f9b\u771f\u6b63\u9ad8\u8d28\u91cf\u3001\u542c\u8d77\u6765\u81ea\u7136\u7684\u8bed\u97f3\uff0c\u5e76\u53ef\u5e7f\u6cdb\u63a7\u5236\u5404\u79cd\u7279\u5f81\uff0c\u5305\u62ec\u6027\u522b\u3001\u80cc\u666f\u566a\u97f3\u3001\u8bed\u901f\u3001\u97f3\u8c03\u548c\u6df7\u54cd\u3002<\/p>\n
\u4f7f\u7528\u5408\u6210\u6807\u6ce8\u7684\u9ad8\u4fdd\u771f\u6587\u672c\u8f6c\u8bed\u97f3\u7684\u81ea\u7136\u8bed\u8a00\u6307\u5bfc\u8bba\u6587\u4ecb\u7ecd\u4e86\u4e00\u79cd\u65b0\u9896\u7684\u6587\u672c\u8f6c\u8bed\u97f3 (TTS) \u7cfb\u7edf\uff0c\u8be5\u7cfb\u7edf\u901a\u8fc7\u5229\u7528\u5927\u89c4\u6a21\u6570\u636e\u96c6\u548c\u81ea\u7136\u8bed\u8a00\u63cf\u8ff0\u5b9e\u73b0\u9ad8\u4fdd\u771f\u548c\u591a\u6837\u5316\u7684\u8bed\u97f3\u751f\u6210\u3002<\/p>\n
\u52a8\u673a\uff1a\u73b0\u6709\u7684 TTS \u7cfb\u7edf\u901a\u5e38\u4f9d\u8d56\u53c2\u8003\u97f3\u9891\u6765\u63a7\u5236\u8bf4\u8bdd\u8005\u7684\u8eab\u4efd\u548c\u98ce\u683c\uff0c\u9650\u5236\u4e86\u5b83\u4eec\u7684\u521b\u9020\u6027\u5e94\u7528\u3002\u867d\u7136\u81ea\u7136\u8bed\u8a00\u63d0\u793a\u63d0\u4f9b\u4e86\u66f4\u76f4\u89c2\u7684\u89e3\u51b3\u65b9\u6848\uff0c\u4f46\u4e4b\u524d\u7684\u5c1d\u8bd5\u53d7\u5230\u7f3a\u4e4f\u5177\u6709\u8bed\u97f3\u5c5e\u6027\u8be6\u7ec6\u81ea\u7136\u8bed\u8a00\u63cf\u8ff0\u7684\u5927\u89c4\u6a21\u6570\u636e\u96c6\u7684\u9650\u5236\u3002<\/p>\n
\u4e3b\u8981\u8d21\u732e\uff1a<\/p>\n
\u6a21\u578b\u67b6\u6784\uff1a<\/p>\n
TTS \u7cfb\u7edf\u91c7\u7528\u4ec5\u89e3\u7801\u5668\u7684 Transformer \u67b6\u6784\u3002\u8f93\u5165\u5305\u62ec\uff1a<\/p>\n
\u89e3\u7801\u5668\u7684\u8f93\u51fa\u662f\u4e00\u7cfb\u5217\u4ee3\u8868\u97f3\u9891\u7279\u5f81\u7684\u79bb\u6563\u6807\u8bb0\uff0c\u7136\u540e\u4f7f\u7528 DAC \u89e3\u7801\u5668\u5c06\u5176\u8f6c\u6362\u4e3a\u97f3\u9891\u6ce2\u5f62\u3002<\/p>\n
\u8bad\u7ec3\uff1a<\/p>\n
\u8be5\u6a21\u578b\u5728\u5927\u89c4\u6a21\u6807\u8bb0\u6570\u636e\u96c6\u4e0a\u8fdb\u884c\u8bad\u7ec3\u3002\u8bad\u7ec3\u8fc7\u7a0b\u6d89\u53ca\u4f18\u5316\u6a21\u578b\u4ee5\u751f\u6210\u4e0e\u63d0\u4f9b\u7684\u6587\u672c\u8bb0\u5f55\u548c\u63cf\u8ff0\u51c6\u786e\u5339\u914d\u7684\u8bed\u97f3\u3002<\/p>\n
\u8bc4\u4f30\uff1a<\/p>\n
\u4f7f\u7528\u5ba2\u89c2\u548c\u4e3b\u89c2\u6307\u6807\u5bf9\u7cfb\u7edf\u8fdb\u884c\u8bc4\u4f30\uff1a<\/p>\n
\u7ed3\u679c\uff1a<\/p>\n
\u7ed3\u679c\u8868\u660e\uff0c\u8be5\u7cfb\u7edf\u53ef\u4ee5\u751f\u6210\u5177\u6709\u5404\u79cd\u53e3\u97f3\u3001\u97f5\u5f8b\u98ce\u683c\u548c\u5f55\u97f3\u6761\u4ef6\u7684\u9ad8\u4fdd\u771f\u8bed\u97f3\uff0c\u7531\u76f4\u89c2\u7684\u81ea\u7136\u8bed\u8a00\u63cf\u8ff0\u63a7\u5236\u3002\u5b83\u5728\u97f3\u9891\u4fdd\u771f\u5ea6\u548c\u63cf\u8ff0\u5339\u914d\u51c6\u786e\u5ea6\u65b9\u9762\u90fd\u4f18\u4e8e\u5e76\u53d1\u5de5\u4f5c\u3002<\/p>\n
\u9996\u5148\uff0c\u5b89\u88c5 Parler-TTS \u5305\uff1a<\/p>\n
! pip install git+https:\/\/github.com\/huggingface\/parler-tts.git<\/code><\/pre>\n\u5e76\u5bfc\u5165\u5fc5\u8981\u7684\u6a21\u5757\uff1a<\/p>\n
import torch\nfrom parler_tts import ParlerTTSForConditionalGeneration, ParlerTTSStreamer\nfrom transformers import AutoTokenizer\nfrom threading import Thread\nimport soundfile as sf<\/code><\/pre>\n\u7136\u540e\u8bbe\u7f6e\u4f60\u559c\u6b22\u7684 torch_device\u3001torch_dtype \u548c model_name\u3002\u6211\u4eec\u8fd8\u4e3a\u586b\u5145\u8bbe\u7f6e\u4e86 max_length \u5e76\u52a0\u8f7d\u9884\u8bad\u7ec3\u6a21\u578b\u548c tokenizer\u3002\u6700\u540e\uff0c\u6211\u4eec\u4ece\u6a21\u578b\u914d\u7f6e\u4e2d\u83b7\u53d6\u91c7\u6837\u7387\u548c\u5e27\u901f\u7387\u3002<\/p>\n
torch_device = \"cuda:0\" # Use \"mps\" for Mac\ntorch_dtype = torch.bfloat16\nmodel_name = \"parler-tts\/parler-tts-mini-v1\"\n\n# need to set padding max length\nmax_length = 50\n\n# load model and tokenizer\ntokenizer = AutoTokenizer.from_pretrained(model_name)\nmodel = ParlerTTSForConditionalGeneration.from_pretrained(\n model_name,\n).to(torch_device, dtype=torch_dtype)\n\nsampling_rate = model.audio_encoder.config.sampling_rate\nframe_rate = model.audio_encoder.config.frame_rate<\/code><\/pre>\n2.1 \u4f7f\u7528\u968f\u673a\u58f0\u97f3\u751f\u6210\u8bed\u97f3<\/h3>\n
\u73b0\u5728\uff0c\u8ba9\u6211\u4eec\u751f\u6210\u4e00\u4e9b\u8bed\u97f3\u3002\u63d0\u4f9b\u6587\u672c\u63d0\u793a\u548c\u6240\u9700\u58f0\u97f3\u548c\u8bf4\u8bdd\u98ce\u683c\u7684\u63cf\u8ff0\u3002\u63cf\u8ff0\u88ab\u6807\u8bb0\u5e76\u8f93\u5165\u5230\u6a21\u578b\u4e2d\uff0c\u6a21\u578b\u4f1a\u751f\u6210\u76f8\u5e94\u7684\u97f3\u9891\u3002\u7136\u540e\uff0c\u4f60\u53ef\u4ee5\u76f4\u63a5\u5728\u7b14\u8bb0\u672c\u4e2d\u6536\u542c\u751f\u6210\u7684\u97f3\u9891\u3002\u70b9\u51fb\u67e5\u770b\u3002<\/p>\n
from IPython.display import Audio\n\nprompt = \"Hey, how are you doing today?\"\ndescription = \"A female speaker delivers a slightly expressive and animated speech with a moderate speed and pitch. The recording is of very high quality, with the speaker's voice sounding clear and very close up.\"\n\ninput_ids = tokenizer(description, return_tensors=\"pt\").input_ids.to(torch_device)\nprompt_input_ids = tokenizer(prompt, return_tensors=\"pt\").input_ids.to(torch_device)\n\ngeneration = model.generate(input_ids=input_ids, prompt_input_ids=prompt_input_ids)\naudio_arr = generation.cpu().float().numpy().squeeze()\nAudio(audio_arr, rate=model.config.sampling_rate)<\/code><\/pre>\n2.2 \u4f7f\u7528\u7279\u5b9a\u8bf4\u8bdd\u8005\u751f\u6210\u8bed\u97f3<\/h3>\n
\u4f60\u8fd8\u53ef\u4ee5\u5f15\u5bfc\u6a21\u578b\u751f\u6210\u7c7b\u4f3c\u4e8e\u7279\u5b9a\u8bf4\u8bdd\u8005\u7684\u8bed\u97f3\u3002\u53ea\u9700\u5728\u63cf\u8ff0\u4e2d\u5305\u542b\u8bf4\u8bdd\u8005\u7684\u59d3\u540d\u6216\u63cf\u8ff0\u6027\u77ed\u8bed\u4ee5\u53ca\u6240\u9700\u7684\u7279\u5f81\u5373\u53ef\u3002<\/p>\n
prompt = \"Hey, how are you doing today?\"\ndescription = \"Jon's voice is monotone yet slightly fast in delivery, with a very close recording that almost has no background noise.\"\n\ninput_ids = tokenizer(description, return_tensors=\"pt\").input_ids.to(torch_device)\nprompt_input_ids = tokenizer(prompt, return_tensors=\"pt\").input_ids.to(torch_device)\n\ngeneration = model.generate(input_ids=input_ids, prompt_input_ids=prompt_input_ids)\naudio_arr = generation.cpu().float().numpy().squeeze()\nAudio(audio_arr, rate=model.config.sampling_rate)<\/code><\/pre>\n\u4f60\u53ef\u4ee5\u4ece 34 \u4e2a\u9884\u5b9a\u4e49\u7684\u8bf4\u8bdd\u8005\u59d3\u540d\u4e2d\u8fdb\u884c\u9009\u62e9\uff0c\u5e76\u6839\u636e\u4f60\u7684\u63cf\u8ff0\u5b9a\u5236\u4ed6\u4eec\u7684\u8bf4\u8bdd\u98ce\u683c\uff08\u4f8b\u5982 Jon\u3001Lea\u3001Gary\u3001Jenna\u3001Mike\u3001Laura\uff09\u3002<\/p>\n
\u63d0\u793a\uff1a<\/p>\n
\n- \u5305\u62ec\u672f\u8bed\u201c\u975e\u5e38\u6e05\u6670\u7684\u97f3\u9891\u201d\u4ee5\u751f\u6210\u6700\u9ad8\u8d28\u91cf\u7684\u97f3\u9891\uff0c\u4ee5\u53ca\u201c\u975e\u5e38\u5608\u6742\u7684\u97f3\u9891\u201d\u4ee5\u751f\u6210\u9ad8\u6c34\u5e73\u7684\u80cc\u666f\u566a\u97f3\u3002<\/li>\n
- \u6807\u70b9\u7b26\u53f7\u53ef\u7528\u4e8e\u63a7\u5236\u751f\u6210\u7684\u97f5\u5f8b\uff0c\u4f8b\u5982\u4f7f\u7528\u9017\u53f7\u5728\u8bed\u97f3\u4e2d\u6dfb\u52a0\u5c0f\u7684\u505c\u987f\u3002<\/li>\n
- \u5176\u4f59\u8bed\u97f3\u7279\u5f81\uff08\u6027\u522b\u3001\u8bed\u901f\u3001\u97f3\u8c03\u548c\u6df7\u54cd\uff09\u53ef\u76f4\u63a5\u901a\u8fc7\u63d0\u793a\u8fdb\u884c\u63a7\u5236\u3002<\/li>\n<\/ul>\n
2.3 \u6d41\u5f0f\u8bed\u97f3\u751f\u6210<\/h3>\n
\u5bf9\u4e8e\u8f83\u957f\u7684\u6587\u672c\u8f93\u5165\uff0c\u4ee5\u5757\u7684\u5f62\u5f0f\u6d41\u5f0f\u4f20\u8f93\u97f3\u9891\u8f93\u51fa\u3002\u8fd9\u5141\u8bb8\u4f60\u5728\u521b\u5efa\u751f\u6210\u7684\u8bed\u97f3\u65f6\u542c\u5230\u5b83\uff0c\u4ece\u800c\u63d0\u4f9b\u66f4\u5177\u4e92\u52a8\u6027\u7684\u4f53\u9a8c\u3002<\/p>\n
def generate(text, description, play_steps_in_s=0.5):\n play_steps = int(frame_rate * play_steps_in_s)\n streamer = ParlerTTSStreamer(model, device=torch_device, play_steps=play_steps)\n # tokenization\n inputs = tokenizer(description, return_tensors=\"pt\").to(torch_device)\n prompt = tokenizer(text, return_tensors=\"pt\").to(torch_device)\n # create generation kwargs\n generation_kwargs = dict(\n input_ids=inputs.input_ids,\n prompt_input_ids=prompt.input_ids,\n attention_mask=inputs.attention_mask,\n prompt_attention_mask=prompt.attention_mask,\n streamer=streamer,\n do_sample=True,\n temperature=1.0,\n min_new_tokens=10,\n )\n # initialize Thread\n thread = Thread(target=model.generate, kwargs=generation_kwargs)\n thread.start()\n # iterate over chunks of audio\n for new_audio in streamer:\n if new_audio.shape[0] == 0:\n break\n yield sampling_rate, new_audio\n\n\n\ntext = \"Parler-TTS is an auto-regressive transformer-based model, meaning generates audio codes (tokens) in a causal fashion.\"\ndescription = \"Mike's talking really fast.\"\n\nchunk_size_in_s = 0.5\n\nfor (sampling_rate, audio_chunk) in generate(text, description, chunk_size_in_s):\n display(Audio(audio_chunk, rate=sampling_rate))<\/code><\/pre>\n3\u3001Parler-TTS \u5fae\u8c03<\/h2>\n
\u672c\u6307\u5357\u6f14\u793a\u4e86\u5728\u5355\u8bf4\u8bdd\u4eba\u6570\u636e\u96c6\u4e0a\u5fae\u8c03 Parler-TTS \u6a21\u578b\u7684\u8fc7\u7a0b\u3002\u6709\u5173\u5b8c\u6574\u7684\u6f14\u7ec3\u548c\u4ee3\u7801\u5b9e\u73b0\uff0c\u8bf7\u53c2\u9605\u63d0\u4f9b\u7684\u5b8c\u6574\u7b14\u8bb0\u672c\u3002<\/p>\n
\u5728\u8fd0\u884c\u7b14\u8bb0\u672c\u4e4b\u524d\uff0c\u8bf7\u901a\u8fc7\u6267\u884c\u4ee5\u4e0b\u547d\u4ee4\u786e\u4fdd\u5df2\u5b89\u88c5\u5fc5\u8981\u7684\u4f9d\u8d56\u9879\uff1a<\/p>\n
\u5b89\u88c5 DataSpeech\uff1a<\/p>\n
!git clone https:\/\/github.com\/huggingface\/dataspeech.git\n!cd dataspeech\n!pip install --quiet -r .\/dataspeech\/requirements.txt<\/code><\/pre>\n\u5b89\u88c5 Parler-TTS\uff1a<\/p>\n
!git clone https:\/\/github.com\/huggingface\/parler-tts.git\n%cd parler-tts\n!pip install --quiet -e .[train]<\/code><\/pre>\n\u5b89\u88c5\u5176\u4ed6\u4f9d\u8d56\u9879\uff1a<\/p>\n
!pip install --upgrade protobuf wandb==0.16.6<\/code><\/pre>\n\u767b\u5f55 Hugging Face\uff1a<\/p>\n
!git config --global credential.helper store\n!huggingface-cli login<\/code><\/pre>\n3.1 \u521b\u5efa\u6211\u4eec\u7684\u5fae\u8c03\u6570\u636e\u96c6<\/h3>\n
\u672c\u8282\u65e8\u5728\u521b\u5efa Jenny TTS \u6570\u636e\u96c6\u7684\u6ce8\u91ca\u7248\u672c\uff0c\u4ee5\u5fae\u8c03 \u3002\u5b8c\u6574\u6570\u636e\u96c6\u53ef\u5728 Hugging Face Hub \u4e0a\u4f5c\u4e3a\u83b7\u53d6\uff0c\u7075\u611f\u6765\u81ea \u3002<\/p>\n
\u51fa\u4e8e\u6f14\u793a\u76ee\u7684\uff0c\u6b64\u7b14\u8bb0\u672c\u4f7f\u75286 \u5c0f\u65f6\u5b50\u96c6\uff1a\u3002\u4f60\u53ef\u4ee5\u901a\u8fc7 Hub \u67e5\u770b\u5668\u6536\u542c\u6837\u672c\u3002<\/p>\n
\u8be5\u8fc7\u7a0b\u5305\u62ec\uff1a1) \u4f7f\u7528\u6d4b\u91cf\u8bed\u97f3\u7279\u5f81\uff08\u8bf4\u8bdd\u901f\u7387\u3001SNR\u3001\u6df7\u54cd\u548c\u8bed\u97f3\u5355\u8c03\u6027\uff09\u7684\u8fde\u7eed\u53d8\u91cf\u6807\u6ce8\u6570\u636e\u96c6\uff0c2) \u5c06\u8fd9\u4e9b\u6ce8\u91ca\u6620\u5c04\u5230\u63cf\u8ff0\u6027\u6587\u672c\u7bb1\uff0c\u4ee5\u53ca 3) \u4ece\u8fd9\u4e9b\u7bb1\u4e2d\u521b\u5efa\u81ea\u7136\u8bed\u8a00\u63cf\u8ff0\u3002<\/p>\n
\u6807\u6ce8\u8fc7\u7a0b\u5229\u7528 \u63d0\u53d6\u8fde\u7eed\u53d8\u91cf\u3002<\/p>\n
\u6709\u5173\u66f4\u8be6\u7ec6\u7684\u8bf4\u660e\uff0c\u8bf7\u53c2\u9605\u3002<\/p>\n
%cd ..\/dataspeech\n!python main.py \"ylacombe\/jenny-tts-6h\" \\ \n --configuration \"default\" \\\n --text_column_name \"transcription\" \\\n --audio_column_name \"audio\" \\\n --cpu_num_workers 2 \\\n --num_workers_per_gpu_for_pitch 2 \\\n --rename_column \\\n --repo_id \"jenny-tts-tags-6h\"<\/code><\/pre>\n3.2 \u5c06\u6807\u6ce8\u6620\u5c04\u5230\u6587\u672c\u7bb1<\/h3>\n
\u4e3a\u4e86\u4e0e \u7684\u8bad\u7ec3\u6570\u636e\u4fdd\u6301\u4e00\u81f4\uff0c\u6211\u4eec\u5c06\u5bf9 Jenny \u6570\u636e\u96c6\u4f7f\u7528\u76f8\u540c\u7684\u6587\u672c\u7bb1\u3002\u4f7f\u7528\u4ee5\u4e0b\u547d\u4ee4\u53ef\u5b9e\u73b0\u6b64\u76ee\u7684\uff1a<\/p>\n
!python .\/scripts\/metadata_to_text.py \\\n \"ylacombe\/jenny-tts-tags-6h\" \\ \n --repo_id \"jenny-tts-tags-6h\" \\\n --configuration \"default\" \\\n --cpu_num_workers 2 \\\n --path_to_bin_edges \".\/examples\/tags_to_annotations\/v01_bin_edges.json\" \\\n --avoid_pitch_computation<\/code><\/pre>\n\u6b64\u811a\u672c\u901a\u8fc7\u5e94\u7528\u6a21\u578b\u521d\u59cb\u8bad\u7ec3\u671f\u95f4\u4f7f\u7528\u7684\u76f8\u540c\u6587\u672c\u5206\u7bb1\u8fc7\u7a0b\u6765\u786e\u4fdd Jenny \u6570\u636e\u96c6\u4e0e Parler-TTS \u6a21\u578b\u4e4b\u95f4\u7684\u517c\u5bb9\u6027\u3002<\/p>\n
3.3 \u4ece\u8fd9\u4e9b\u6587\u672c\u7bb1\u521b\u5efa\u81ea\u7136\u8bed\u8a00\u63cf\u8ff0<\/h3>\n
\u5c06\u6587\u672c\u7bb1\u5206\u914d\u7ed9 Jenny \u6570\u636e\u96c6\u540e\uff0c\u4e0b\u4e00\u6b65\u662f\u6839\u636e\u8fd9\u4e9b\u7279\u5f81\u751f\u6210\u81ea\u7136\u8bed\u8a00\u63cf\u8ff0\u3002\u6211\u4eec\u9009\u62e9\u7684\u65b9\u6cd5\u662f\u521b\u5efa\u5305\u542b\u540d\u5b57\u201cJenny\u201d\u5e76\u63cf\u8ff0\u8bed\u97f3\u7279\u5f81\u7684\u63d0\u793a\uff0c\u4ece\u800c\u5f97\u5230\u5982\u4e0b\u63d0\u793a\uff1a\u201cJenny \u7684\u53d1\u97f3\u975e\u5e38\u6709\u8868\u73b0\u529b\uff0c\u4f46\u53d1\u97f3\u975e\u5e38\u6162\u3002\u623f\u95f4\u91cc\u6709\u4e00\u4e9b\u80cc\u666f\u566a\u97f3\uff0c\u8fd8\u6709\u4e00\u70b9\u56de\u58f0\u3002\u201d<\/p>\n
\u8fd9\u4e2a\u8d44\u6e90\u5bc6\u96c6\u578b\u6b65\u9aa4\u901a\u5e38\u9700\u8981\u5229\u7528 GPU\u3002\u4ee5\u4e0b\u547d\u4ee4\u4f7f\u7528 Google \u7684 \u6f14\u793a\u4e86\u6b64\u8fc7\u7a0b\uff0c\u8be5\u8fc7\u7a0b\u5728 Colab \u514d\u8d39 T4 \u4e0a\u5927\u7ea6\u9700\u8981 15 \u5206\u949f\u5b8c\u6210\u3002<\/p>\n
!python .\/scripts\/run_prompt_creation.py \\\n --speaker_name \"Jenny\" \\\n --is_single_speaker \\\n --dataset_name \"ylacombe\/jenny-tts-tags-6h\" \\\n --output_dir \".\/tmp_jenny\" \\\n --dataset_config_name \"default\" \\\n --model_name_or_path \"google\/gemma-2b-it\" \\\n --per_device_eval_batch_size 12 \\\n --attn_implementation \"sdpa\" \\\n --dataloader_num_workers 2 \\\n --push_to_hub \\\n --hub_dataset_id \"jenny-tts-6h-tagged\" \\\n --preprocessing_num_workers 2<\/code><\/pre>\n\u8be5\u547d\u4ee4\u6307\u5b9a\u8981\u6807\u6ce8\u7684\u6570\u636e\u96c6\u548c\u914d\u7f6e\u3002 model_name_or_path<\/code> \u53c2\u6570\u6307\u5411\u9002\u5408\u8fdb\u884c\u5feb\u901f\u6807\u6ce8\u7684 transformers \u6a21\u578b\uff0c\u6b64\u5904\u63d0\u4f9b\u4e86\u4e00\u7cfb\u5217\u9009\u9879\u3002<\/p>\n\u503c\u5f97\u6ce8\u610f\u7684\u662f\uff0c\u6807\u5fd7 --speaker_name \"Jenny\" --is_single_speaker<\/code> \u660e\u786e\u5c06\u6570\u636e\u96c6\u5b9a\u4e49\u4e3a\u5355\u58f0\u9053\u8bf4\u8bdd\u8005\uff0c\u5e76\u5c06\u8bed\u97f3\u6807\u8bc6\u4e3a\u201cJenny\u201d\u3002<\/p>\n3.4 \u59cb\u5fae\u8c03 Parler-TTS<\/h3>\n
\u51c6\u5907\u597d\u5e26\u6807\u6ce8\u7684 Jenny \u6570\u636e\u96c6\u540e\uff0c\u6211\u4eec\u73b0\u5728\u53ef\u4ee5\u4e13\u6ce8\u4e8e\u5fae\u8c03 Parler-TTS \u6a21\u578b\u3002\u5728\u63d0\u4f9b\u4e86\u4e00\u4e2a\u65b9\u4fbf\u7684\u8bad\u7ec3\u811a\u672c\uff0c\u53ef\u7b80\u5316\u6b64\u8fc7\u7a0b\u3002<\/p>\n
\n \u6ce8\u610f\uff1a\u811a\u672c\u9700\u8981\u505a\u51fa\u6709\u5173\u6743\u91cd\u548c\u504f\u5dee (WandB) \u65e5\u5fd7\u8bb0\u5f55\u7684\u51b3\u5b9a\u3002\u5982\u679c\u4f60\u6ca1\u6709\u5e10\u6237\uff0c\u8bf7\u9009\u62e9\u9009\u9879 3 \u4ee5\u7981\u7528\u65e5\u5fd7\u8bb0\u5f55\u3002\u5426\u5219\uff0c\u8bf7\u542f\u7528\u65e5\u5fd7\u8bb0\u5f55\u4ee5\u8ddf\u8e2a\u6a21\u578b\u7684\u8bad\u7ec3\u8fdb\u5ea6\u3002\n<\/p><\/blockquote>\n
\u4ee5\u4e0b\u547d\u4ee4\u542f\u52a8\u5fae\u8c03\u8fc7\u7a0b\uff1a<\/p>\n
%cd ..\/parler-tts\n\n!accelerate launch .\/training\/run_parler_tts_training.py \\\n --model_name_or_path \"parler-tts\/parler_tts_mini_v0.1\" \\\n --feature_extractor_name \"parler-tts\/dac_44khZ_8kbps\" \\\n --description_tokenizer_name \"parler-tts\/parler_tts_mini_v0.1\" \\\n --prompt_tokenizer_name \"parler-tts\/parler_tts_mini_v0.1\" \\\n --report_to \"wandb\" \\\n --overwrite_output_dir true \\\n --train_dataset_name \"ylacombe\/jenny-tts-6h\" \\\n --train_metadata_dataset_name \"ylacombe\/jenny-tts-6h-tagged\" \\\n --train_dataset_config_name \"default\" \\\n --train_split_name \"train\" \\\n --eval_dataset_name \"ylacombe\/jenny-tts-6h\" \\\n --eval_metadata_dataset_name \"ylacombe\/jenny-tts-6h-tagged\" \\\n --eval_dataset_config_name \"default\" \\\n --eval_split_name \"train\" \\\n --max_eval_samples 8 \\\n --per_device_eval_batch_size 8 \\\n --target_audio_column_name \"audio\" \\\n --description_column_name \"text_description\" \\\n --prompt_column_name \"text\" \\\n --max_duration_in_seconds 20 \\\n --min_duration_in_seconds 2.0 \\\n --max_text_length 400 \\\n --preprocessing_num_workers 2 \\\n --do_train true \\\n --num_train_epochs 2 \\\n --gradient_accumulation_steps 18 \\\n --gradient_checkpointing true \\\n --per_device_train_batch_size 2 \\\n --learning_rate 0.00008 \\\n --adam_beta1 0.9 \\\n --adam_beta2 0.99 \\\n --weight_decay 0.01 \\\n --lr_scheduler_type \"constant_with_warmup\" \\\n --warmup_steps 50 \\\n --logging_steps 2 \\\n --freeze_text_encoder true \\\n --audio_encoder_per_device_batch_size 4 \\\n --dtype \"float16\" \\\n --seed 456 \\\n --output_dir \".\/output_dir_training\/\" \\\n --temporary_save_to_disk \".\/audio_code_tmp\/\" \\\n --save_to_disk \".\/tmp_dataset_audio\/\" \\\n --dataloader_num_workers 2 \\\n --do_eval \\\n --predict_with_generate \\\n --include_inputs_for_metrics \\\n --group_by_length true<\/code><\/pre>\n\u6b64\u811a\u672c\u5229\u7528\u5e26\u6807\u6ce8\u7684 Jenny \u6570\u636e\u96c6\u548c\u6307\u5b9a\u7684\u53c2\u6570\u5bf9 Parler-TTS \u6a21\u578b\u8fdb\u884c\u5fae\u8c03\uff0c\u6700\u7ec8\u63d0\u9ad8\u5176\u6027\u80fd\u5e76\u4f7f\u5176\u9002\u5e94 Jenny \u8bed\u97f3\u7684\u7279\u70b9\u3002<\/p>\n
\n\n","protected":false},"excerpt":{"rendered":"
Parler-TTS \u9879\u76ee\u5f88\u9ad8\u5174\u5730\u5ba3\u5e03\u53d1\u5e03\u4e24\u4e2a\u65b0\u7684\u6587\u672c\u8f6c\u8bed\u97f3\u6a21\u578b\uff01\u672c\u6587\u4ecb\u7ecd\u5982\u4f55\u4f7f\u7528Gemma\u5236\u4f5c\u81ea\u5df1\u7684\u6570\u636e\u96c6\uff0c\u5982\u4f55\u5fae\u8c03Parler-TTS\uff0c\u4ee5\u53ca\u5982\u4f55\u4f7f\u7528Parler-TTS\u6a21\u578b\u8fdb\u884c\u63a8\u7406\u3002 1\u3001Parler-TTS \u6a21\u578b\u7b80\u4ecb \u9996\u5148\uff0c\u6211\u4eec\u6709 \uff0c\u8fd9\u662f\u4e00\u6b3e\u8f7b\u91cf\u7ea7\u6a21\u578b\uff0c\u975e\u5e38\u9002\u5408\u5feb\u901f\u8f7b\u677e\u5730\u751f\u6210\u8bed\u97f3\u3002\u53d7\u6700\u8fd1\u7814\u7a76\u8bba\u6587\u300a\u4f7f\u7528\u5408\u6210\u6807\u6ce8\u7684\u9ad8\u4fdd\u771f\u6587\u672c\u8f6c\u8bed\u97f3\u7684\u81ea\u7136\u8bed\u8a00\u6307\u5bfc\u300b\u7684\u542f\u53d1\uff0cParler-TTS Mini v0.1 \u8ba9\u4f60\u901a\u8fc7\u7b80\u5355\u7684\u6587\u672c\u63d0\u793a\u76f4\u89c2\u5730\u63a7\u5236\u5404\u79cd\u8bed\u97f3\u65b9\u9762\uff0c\u4f8b\u5982\u6027\u522b\u3001\u80cc\u666f\u566a\u97f3\u548c\u8bed\u901f\u3002 \u5bf9\u4e8e\u90a3\u4e9b\u5bfb\u6c42\u6700\u5927\u8868\u73b0\u529b\u548c\u63a7\u5236\u529b\u7684\u4eba\uff0c\u6211\u4eec\u8fd8\u6709 \u3002\u8fd9\u4e2a 2.2B \u53c2\u6570\u6a21\u578b\u7ecf\u8fc7\u5927\u91cf 45K \u5c0f\u65f6\u7684\u97f3\u9891\u6570\u636e\u8bad\u7ec3\uff0c\u53ef\u63d0\u4f9b\u771f\u6b63\u9ad8\u8d28\u91cf\u3001\u542c\u8d77\u6765\u81ea\u7136\u7684\u8bed\u97f3\uff0c\u5e76\u53ef\u5e7f\u6cdb\u63a7\u5236\u5404\u79cd\u7279\u5f81\uff0c\u5305\u62ec\u6027\u522b\u3001\u80cc\u666f\u566a\u97f3\u3001\u8bed\u901f\u3001\u97f3\u8c03\u548c\u6df7\u54cd\u3002 \u4f7f\u7528\u5408\u6210\u6807\u6ce8\u7684\u9ad8\u4fdd\u771f\u6587\u672c\u8f6c\u8bed\u97f3\u7684\u81ea\u7136\u8bed\u8a00\u6307\u5bfc\u8bba\u6587\u4ecb\u7ecd\u4e86\u4e00\u79cd\u65b0\u9896\u7684\u6587\u672c\u8f6c\u8bed\u97f3 (TTS) \u7cfb\u7edf\uff0c\u8be5\u7cfb\u7edf\u901a\u8fc7\u5229\u7528\u5927\u89c4\u6a21\u6570\u636e\u96c6\u548c\u81ea\u7136\u8bed\u8a00\u63cf\u8ff0\u5b9e\u73b0\u9ad8\u4fdd\u771f\u548c\u591a\u6837\u5316\u7684\u8bed\u97f3\u751f\u6210\u3002 \u52a8\u673a\uff1a\u73b0\u6709\u7684 TTS \u7cfb\u7edf\u901a\u5e38\u4f9d\u8d56\u53c2\u8003\u97f3\u9891\u6765\u63a7\u5236\u8bf4\u8bdd\u8005\u7684\u8eab\u4efd\u548c\u98ce\u683c\uff0c\u9650\u5236\u4e86\u5b83\u4eec\u7684\u521b\u9020\u6027\u5e94\u7528\u3002\u867d\u7136\u81ea\u7136\u8bed\u8a00\u63d0\u793a\u63d0\u4f9b\u4e86\u66f4\u76f4\u89c2\u7684\u89e3\u51b3\u65b9\u6848\uff0c\u4f46\u4e4b\u524d\u7684\u5c1d\u8bd5\u53d7\u5230\u7f3a\u4e4f\u5177\u6709\u8bed\u97f3\u5c5e\u6027\u8be6\u7ec6\u81ea\u7136\u8bed\u8a00\u63cf\u8ff0\u7684\u5927\u89c4\u6a21\u6570\u636e\u96c6\u7684\u9650\u5236\u3002 \u4e3b\u8981\u8d21\u732e\uff1a \u81ea\u52a8\u6807\u8bb0\uff1a\u4f5c\u8005\u901a\u8fc7\u63d0\u51fa\u4e00\u79cd\u53ef\u6269\u5c55\u7684\u65b9\u6cd5\u6765\u89e3\u51b3\u6570\u636e\u7a00\u7f3a\u95ee\u9898\uff0c\u8be5\u65b9\u6cd5\u7528\u4e8e\u81ea\u52a8\u6807\u8bb0\u5927\u91cf 45k \u5c0f\u65f6\u7684\u8bed\u97f3\u6570\u636e\u96c6 (Multilingual LibriSpeech)\uff0c\u5176\u4e2d\u5305\u542b\u6027\u522b\u3001\u53e3\u97f3\u3001\u8bed\u901f\u3001\u97f3\u8c03\u548c\u5f55\u97f3\u8d28\u91cf\u7b49\u5404\u79cd\u5c5e\u6027\u3002\u5b83\u4eec\u8fd8\u5305\u62ec\u4e00\u4e2a\u8f83\u5c0f\u7684\u9ad8\u4fdd\u771f\u6570\u636e\u96c6 (LibriTTS-R)\uff0c\u4ee5\u63d0\u9ad8\u97f3\u9891\u8d28\u91cf\u3002 \u8bed\u97f3\u8bed\u8a00\u6a21\u578b\uff1a\u6b64\u6807\u8bb0\u6570\u636e\u96c6\u7528\u4e8e\u8bad\u7ec3\u6539\u7f16\u81ea\u901a\u7528\u97f3\u9891\u751f\u6210\u5e93 AudioCraft \u7684\u8bed\u97f3\u8bed\u8a00\u6a21\u578b\u3002\u6a21\u578b\u67b6\u6784\u4f7f\u7528\u4ec5\u89e3\u7801\u5668\u7684 Transformer \u7f51\u7edc\u3002\u4e0e\u4ee5\u524d\u7684\u5de5\u4f5c\u4e0d\u540c\uff0c\u8be5\u6a21\u578b\u4ec5\u4f9d\u9760\u81ea\u7136\u8bed\u8a00\u63cf\u8ff0\u6765\u63a7\u5236\u8bed\u97f3\u5c5e\u6027\uff0c\u6ca1\u6709\u4efb\u4f55\u53c2\u8003\u97f3\u9891\u5d4c\u5165\u3002 \u9ad8\u4fdd\u771f\u97f3\u9891\uff1a\u8be5\u7cfb\u7edf\u5229\u7528 Descript Audio Codec (DAC) \u751f\u6210\u9ad8\u4fdd\u771f\u97f3\u9891\uff0c\u5373\u4f7f\u5728\u6709\u9650\u7684\u9ad8\u4fdd\u771f\u8bad\u7ec3\u6570\u636e\u4e0b\uff0c\u4e5f\u80fd\u6bd4\u4ee5\u524d\u7684\u65b9\u6cd5\u53d6\u5f97\u663e\u7740\u6539\u8fdb\u3002 \u6a21\u578b\u67b6\u6784\uff1a TTS \u7cfb\u7edf\u91c7\u7528\u4ec5\u89e3\u7801\u5668\u7684 Transformer \u67b6\u6784\u3002\u8f93\u5165\u5305\u62ec\uff1a \u8f6c\u5f55\u6587\u672c\uff1a\u8981\u5408\u6210\u7684\u6587\u672c\uff0c\u9644\u52a0\u5728\u8f93\u5165\u5e8f\u5217\u7684\u524d\u9762\u3002 \u63cf\u8ff0\u6587\u672c\uff1a\u6240\u9700\u8bed\u97f3\u7279\u5f81\u7684\u81ea\u7136\u8bed\u8a00\u63cf\u8ff0\uff08\u4f8b\u5982\uff0c\u201c\u4e00\u4e2a\u5e26\u6709\u82f1\u56fd\u53e3\u97f3\u7684\u7537\u4eba\u7528\u6d3b\u6cfc\u7684\u8bed\u8c03\u8bf4\u8bdd\u201d\uff09\u3002\u8fd9\u7531\u9884\u5148\u8bad\u7ec3\u7684 T5 \u6587\u672c\u7f16\u7801\u5668\u5904\u7406\uff0c\u5e76\u901a\u8fc7\u4ea4\u53c9\u6ce8\u610f\u9988\u9001\u5230\u89e3\u7801\u5668\u3002 \u89e3\u7801\u5668\u7684\u8f93\u51fa\u662f\u4e00\u7cfb\u5217\u4ee3\u8868\u97f3\u9891\u7279\u5f81\u7684\u79bb\u6563\u6807\u8bb0\uff0c\u7136\u540e\u4f7f\u7528 DAC \u89e3\u7801\u5668\u5c06\u5176\u8f6c\u6362\u4e3a\u97f3\u9891\u6ce2\u5f62\u3002 \u8bad\u7ec3\uff1a \u8be5\u6a21\u578b\u5728\u5927\u89c4\u6a21\u6807\u8bb0\u6570\u636e\u96c6\u4e0a\u8fdb\u884c\u8bad\u7ec3\u3002\u8bad\u7ec3\u8fc7\u7a0b\u6d89\u53ca\u4f18\u5316\u6a21\u578b\u4ee5\u751f\u6210\u4e0e\u63d0\u4f9b\u7684\u6587\u672c\u8bb0\u5f55\u548c\u63cf\u8ff0\u51c6\u786e\u5339\u914d\u7684\u8bed\u97f3\u3002 \u8bc4\u4f30\uff1a \u4f7f\u7528\u5ba2\u89c2\u548c\u4e3b\u89c2\u6307\u6807\u5bf9\u7cfb\u7edf\u8fdb\u884c\u8bc4\u4f30\uff1a […]<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"closed","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[13],"tags":[],"class_list":["post-53763","post","type-post","status-publish","format-standard","hentry","category-ai"],"_links":{"self":[{"href":"https:\/\/fwq.ai\/blog\/wp-json\/wp\/v2\/posts\/53763","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/fwq.ai\/blog\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/fwq.ai\/blog\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/fwq.ai\/blog\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/fwq.ai\/blog\/wp-json\/wp\/v2\/comments?post=53763"}],"version-history":[{"count":0,"href":"https:\/\/fwq.ai\/blog\/wp-json\/wp\/v2\/posts\/53763\/revisions"}],"wp:attachment":[{"href":"https:\/\/fwq.ai\/blog\/wp-json\/wp\/v2\/media?parent=53763"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/fwq.ai\/blog\/wp-json\/wp\/v2\/categories?post=53763"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/fwq.ai\/blog\/wp-json\/wp\/v2\/tags?post=53763"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}