Voice tone cloning with OpenVoice and OpenVINO

This Jupyter notebook can be launched on-line, opening an interactive environment in a browser window. You can also make a local installation. Choose one of the following options:

OpenVoice is a versatile instant voice tone transferring and generating speech in various languages with just a brief audio snippet from the source speaker. OpenVoice has three main features: (i) high quality tone color replication with multiple languages and accents; (ii) it provides fine-tuned control over voice styles, including emotions, accents, as well as other parameters such as rhythm, pauses, and intonation. (iii) OpenVoice achieves zero-shot cross-lingual voice cloning, eliminating the need for the generated speech and the reference speech to be part of a massive-speaker multilingual training dataset.

image

More details about model can be found in project web page, paper, and official repository

This notebook provides example of converting PyTorch OpenVoice model to OpenVINO IR. In this tutorial we will explore how to convert and run OpenVoice using OpenVINO.

Table of contents:

• Clone repository and install requirements

• Download checkpoints and load PyTorch model

• Convert models to OpenVINO IR

• Inference

  • Select inference device

  • Select reference tone

  • Run inference

• Run OpenVoice Gradio interactive demo

• Cleanup

Installation Instructions

This is a self-contained example that relies solely on its own code.

We recommend running the notebook in a virtual environment. You only need a Jupyter server to start. For details, please refer to Installation Guide.

Clone repository and install requirements

# Fetch `notebook_utils` module
import requests

r = requests.get(
    url="https://raw.githubusercontent.com/openvinotoolkit/openvino_notebooks/latest/utils/notebook_utils.py",
)
open("notebook_utils.py", "w").write(r.text)

r = requests.get(
    url="https://raw.githubusercontent.com/openvinotoolkit/openvino_notebooks/latest/utils/cmd_helper.py",
)
open("cmd_helper.py", "w").write(r.text)


from notebook_utils import download_file, device_widget

from pathlib import Path

from cmd_helper import clone_repo


repo_dir = Path("OpenVoice")

if not repo_dir.exists():
    clone_repo("https://github.com/myshell-ai/OpenVoice")
    orig_english_path = Path("OpenVoice/openvoice/text/_orig_english.py")
    english_path = Path("OpenVoice/openvoice/text/english.py")

    english_path.rename(orig_english_path)

    with orig_english_path.open("r") as f:
        data = f.read()
        data = data.replace("unidecode", "anyascii")
        with english_path.open("w") as out_f:
            out_f.write(data)


# fix a problem with silero downloading and installing
with Path("OpenVoice/openvoice/se_extractor.py").open("r") as orig_file:
    data = orig_file.read()
    data = data.replace("method=\"silero\"", "method=\"silero:3.0\"")
    with Path("OpenVoice/openvoice/se_extractor.py").open("w") as out_f:
            out_f.write(data)


%pip install -q "librosa>=0.8.1" "wavmark>=0.0.3" "faster-whisper>=0.9.0" "pydub>=0.25.1" "whisper-timestamped>=1.14.2" "tqdm" "inflect>=7.0.0" "eng_to_ipa>=0.0.2" "pypinyin>=0.50.0" \
"cn2an>=0.5.22" "jieba>=0.42.1" "langid>=1.1.6" "gradio>=4.15" "ipywebrtc" "anyascii" "openvino>=2023.3" "torch>=2.1" "nncf>=2.11.0"

ERROR: pip's dependency resolver does not currently take into account all the packages that are installed. This behaviour is the source of the following dependency conflicts.
tensorflow 2.12.0 requires keras<2.13,>=2.12.0, but you have keras 2.13.1 which is incompatible.
tensorflow 2.12.0 requires numpy<1.24,>=1.22, but you have numpy 1.24.4 which is incompatible.
tensorflow 2.12.0 requires tensorboard<2.13,>=2.12, but you have tensorboard 2.13.0 which is incompatible.
tensorflow 2.12.0 requires tensorflow-estimator<2.13,>=2.12.0, but you have tensorflow-estimator 2.13.0 which is incompatible.
tensorflow-cpu 2.13.1 requires numpy<=1.24.3,>=1.22, but you have numpy 1.24.4 which is incompatible.
tensorflow-cpu 2.13.1 requires typing-extensions<4.6.0,>=3.6.6, but you have typing-extensions 4.12.2 which is incompatible.
Note: you may need to restart the kernel to use updated packages.

Download checkpoints and load PyTorch model

import os
import torch
import openvino as ov
import ipywidgets as widgets
from IPython.display import Audio

core = ov.Core()

from openvoice.api import BaseSpeakerTTS, ToneColorConverter, OpenVoiceBaseClass
import openvoice.se_extractor as se_extractor

Importing the dtw module. When using in academic works please cite:
  T. Giorgino. Computing and Visualizing Dynamic Time Warping Alignments in R: The dtw Package.
  J. Stat. Soft., doi:10.18637/jss.v031.i07.

CKPT_BASE_PATH = Path("checkpoints")

en_suffix = CKPT_BASE_PATH / "base_speakers/EN"
zh_suffix = CKPT_BASE_PATH / "base_speakers/ZH"
converter_suffix = CKPT_BASE_PATH / "converter"

To make notebook lightweight by default model for Chinese speech is not activated, in order turn on please set flag enable_chinese_lang to True

enable_chinese_lang = False

def download_from_hf_hub(filename, local_dir="./"):
    from huggingface_hub import hf_hub_download

    local_path = Path(local_dir)
    local_path.mkdir(exist_ok=True)
    hf_hub_download(repo_id="myshell-ai/OpenVoice", filename=filename, local_dir=local_path)


download_from_hf_hub(f"{converter_suffix.as_posix()}/checkpoint.pth")
download_from_hf_hub(f"{converter_suffix.as_posix()}/config.json")
download_from_hf_hub(f"{en_suffix.as_posix()}/checkpoint.pth")
download_from_hf_hub(f"{en_suffix.as_posix()}/config.json")

download_from_hf_hub(f"{en_suffix.as_posix()}/en_default_se.pth")
download_from_hf_hub(f"{en_suffix.as_posix()}/en_style_se.pth")

if enable_chinese_lang:
    download_from_hf_hub(f"{zh_suffix.as_posix()}/checkpoint.pth")
    download_from_hf_hub(f"{zh_suffix.as_posix()}/config.json")
    download_from_hf_hub(f"{zh_suffix.as_posix()}/zh_default_se.pth")

checkpoint.pth:   0%|          | 0.00/131M [00:00<?, ?B/s]

checkpoints/converter/config.json:   0%|          | 0.00/850 [00:00<?, ?B/s]

checkpoint.pth:   0%|          | 0.00/160M [00:00<?, ?B/s]

checkpoints/base_speakers/EN/config.json:   0%|          | 0.00/1.97k [00:00<?, ?B/s]

en_default_se.pth:   0%|          | 0.00/1.79k [00:00<?, ?B/s]

en_style_se.pth:   0%|          | 0.00/1.78k [00:00<?, ?B/s]

pt_device = "cpu"

en_base_speaker_tts = BaseSpeakerTTS(en_suffix / "config.json", device=pt_device)
en_base_speaker_tts.load_ckpt(en_suffix / "checkpoint.pth")

tone_color_converter = ToneColorConverter(converter_suffix / "config.json", device=pt_device)
tone_color_converter.load_ckpt(converter_suffix / "checkpoint.pth")

if enable_chinese_lang:
    zh_base_speaker_tts = BaseSpeakerTTS(zh_suffix / "config.json", device=pt_device)
    zh_base_speaker_tts.load_ckpt(zh_suffix / "checkpoint.pth")
else:
    zh_base_speaker_tts = None

/opt/home/k8sworker/ci-ai/cibuilds/jobs/ov-notebook/jobs/OVNotebookOps/builds/835/archive/.workspace/scm/ov-notebook/.venv/lib/python3.8/site-packages/torch/nn/utils/weight_norm.py:134: FutureWarning: torch.nn.utils.weight_norm is deprecated in favor of torch.nn.utils.parametrizations.weight_norm.
  WeightNorm.apply(module, name, dim)
/opt/home/k8sworker/ci-ai/cibuilds/jobs/ov-notebook/jobs/OVNotebookOps/builds/835/archive/.workspace/scm/ov-notebook/notebooks/openvoice/OpenVoice/openvoice/api.py:36: FutureWarning: You are using torch.load with weights_only=False (the current default value), which uses the default pickle module implicitly. It is possible to construct malicious pickle data which will execute arbitrary code during unpickling (See pytorch/pytorch for more details). In a future release, the default value for weights_only will be flipped to True. This limits the functions that could be executed during unpickling. Arbitrary objects will no longer be allowed to be loaded via this mode unless they are explicitly allowlisted by the user via torch.serialization.add_safe_globals. We recommend you start setting weights_only=True for any use case where you don't have full control of the loaded file. Please open an issue on GitHub for any issues related to this experimental feature.
  checkpoint_dict = torch.load(ckpt_path, map_location=torch.device(self.device))

Loaded checkpoint 'checkpoints/base_speakers/EN/checkpoint.pth'
missing/unexpected keys: [] []
Loaded checkpoint 'checkpoints/converter/checkpoint.pth'
missing/unexpected keys: [] []

/opt/home/k8sworker/ci-ai/cibuilds/jobs/ov-notebook/jobs/OVNotebookOps/builds/835/archive/.workspace/scm/ov-notebook/.venv/lib/python3.8/site-packages/wavmark/__init__.py:16: FutureWarning: You are using torch.load with weights_only=False (the current default value), which uses the default pickle module implicitly. It is possible to construct malicious pickle data which will execute arbitrary code during unpickling (See pytorch/pytorch for more details). In a future release, the default value for weights_only will be flipped to True. This limits the functions that could be executed during unpickling. Arbitrary objects will no longer be allowed to be loaded via this mode unless they are explicitly allowlisted by the user via torch.serialization.add_safe_globals. We recommend you start setting weights_only=True for any use case where you don't have full control of the loaded file. Please open an issue on GitHub for any issues related to this experimental feature.
  checkpoint = torch.load(resume_path, map_location=torch.device('cpu'))
/opt/home/k8sworker/ci-ai/cibuilds/jobs/ov-notebook/jobs/OVNotebookOps/builds/835/archive/.workspace/scm/ov-notebook/notebooks/openvoice/OpenVoice/openvoice/api.py:36: FutureWarning: You are using torch.load with weights_only=False (the current default value), which uses the default pickle module implicitly. It is possible to construct malicious pickle data which will execute arbitrary code during unpickling (See pytorch/pytorch for more details). In a future release, the default value for weights_only will be flipped to True. This limits the functions that could be executed during unpickling. Arbitrary objects will no longer be allowed to be loaded via this mode unless they are explicitly allowlisted by the user via torch.serialization.add_safe_globals. We recommend you start setting weights_only=True for any use case where you don't have full control of the loaded file. Please open an issue on GitHub for any issues related to this experimental feature.
  checkpoint_dict = torch.load(ckpt_path, map_location=torch.device(self.device))

Convert models to OpenVINO IR

There are 2 models in OpenVoice: first one is responsible for speech generation BaseSpeakerTTS and the second one ToneColorConverter imposes arbitrary voice tone to the original speech. To convert to OpenVino IR format first we need to get acceptable torch.nn.Module object. Both ToneColorConverter, BaseSpeakerTTS instead of using self.forward as the main entry point use custom infer and convert_voice methods respectively, therefore need to wrap them with a custom class that is inherited from torch.nn.Module.

class OVOpenVoiceBase(torch.nn.Module):
    """
    Base class for both TTS and voice tone conversion model: constructor is same for both of them.
    """

    def __init__(self, voice_model: OpenVoiceBaseClass):
        super().__init__()
        self.voice_model = voice_model
        for par in voice_model.model.parameters():
            par.requires_grad = False


class OVOpenVoiceTTS(OVOpenVoiceBase):
    """
    Constructor of this class accepts BaseSpeakerTTS object for speech generation and wraps it's 'infer' method with forward.
    """

    def get_example_input(self):
        stn_tst = self.voice_model.get_text("this is original text", self.voice_model.hps, False)
        x_tst = stn_tst.unsqueeze(0)
        x_tst_lengths = torch.LongTensor([stn_tst.size(0)])
        speaker_id = torch.LongTensor([1])
        noise_scale = torch.tensor(0.667)
        length_scale = torch.tensor(1.0)
        noise_scale_w = torch.tensor(0.6)
        return (
            x_tst,
            x_tst_lengths,
            speaker_id,
            noise_scale,
            length_scale,
            noise_scale_w,
        )

    def forward(self, x, x_lengths, sid, noise_scale, length_scale, noise_scale_w):
        return self.voice_model.model.infer(x, x_lengths, sid, noise_scale, length_scale, noise_scale_w)


class OVOpenVoiceConverter(OVOpenVoiceBase):
    """
    Constructor of this class accepts ToneColorConverter object for voice tone conversion and wraps it's 'voice_conversion' method with forward.
    """

    def get_example_input(self):
        y = torch.randn([1, 513, 238], dtype=torch.float32)
        y_lengths = torch.LongTensor([y.size(-1)])
        target_se = torch.randn(*(1, 256, 1))
        source_se = torch.randn(*(1, 256, 1))
        tau = torch.tensor(0.3)
        return (y, y_lengths, source_se, target_se, tau)

    def forward(self, y, y_lengths, sid_src, sid_tgt, tau):
        return self.voice_model.model.voice_conversion(y, y_lengths, sid_src, sid_tgt, tau)

Convert to OpenVino IR and save to IRs_path folder for the future use. If IRs already exist skip conversion and read them directly

For reducing memory consumption, weights compression optimization can be applied using NNCF. Weight compression aims to reduce the memory footprint of a model. models, which require extensive memory to store the weights during inference, can benefit from weight compression in the following ways:

• enabling the inference of exceptionally large models that cannot be accommodated in the memory of the device;

• improving the inference performance of the models by reducing the latency of the memory access when computing the operations with weights, for example, Linear layers.

Neural Network Compression Framework (NNCF) provides 4-bit / 8-bit mixed weight quantization as a compression method. The main difference between weights compression and full model quantization (post-training quantization) is that activations remain floating-point in the case of weights compression which leads to a better accuracy. In addition, weight compression is data-free and does not require a calibration dataset, making it easy to use.

nncf.compress_weights function can be used for performing weights compression. The function accepts an OpenVINO model and other compression parameters.

More details about weights compression can be found in OpenVINO documentation.

import nncf


IRS_PATH = Path("openvino_irs/")
EN_TTS_IR = IRS_PATH / "openvoice_en_tts.xml"
ZH_TTS_IR = IRS_PATH / "openvoice_zh_tts.xml"
VOICE_CONVERTER_IR = IRS_PATH / "openvoice_tone_conversion.xml"

paths = [EN_TTS_IR, VOICE_CONVERTER_IR]
models = [
    OVOpenVoiceTTS(en_base_speaker_tts),
    OVOpenVoiceConverter(tone_color_converter),
]
if enable_chinese_lang:
    models.append(OVOpenVoiceTTS(zh_base_speaker_tts))
    paths.append(ZH_TTS_IR)
ov_models = []

for model, path in zip(models, paths):
    if not path.exists():
        ov_model = ov.convert_model(model, example_input=model.get_example_input())
        ov_model = nncf.compress_weights(ov_model)
        ov.save_model(ov_model, path)
    else:
        ov_model = core.read_model(path)
    ov_models.append(ov_model)

ov_en_tts, ov_voice_conversion = ov_models[:2]
if enable_chinese_lang:
    ov_zh_tts = ov_models[-1]

INFO:nncf:NNCF initialized successfully. Supported frameworks detected: torch, tensorflow, onnx, openvino
this is original text.
 length:22
 length:21

/opt/home/k8sworker/ci-ai/cibuilds/jobs/ov-notebook/jobs/OVNotebookOps/builds/835/archive/.workspace/scm/ov-notebook/notebooks/openvoice/OpenVoice/openvoice/attentions.py:283: TracerWarning: Converting a tensor to a Python boolean might cause the trace to be incorrect. We can't record the data flow of Python values, so this value will be treated as a constant in the future. This means that the trace might not generalize to other inputs!
  assert (
/opt/home/k8sworker/ci-ai/cibuilds/jobs/ov-notebook/jobs/OVNotebookOps/builds/835/archive/.workspace/scm/ov-notebook/notebooks/openvoice/OpenVoice/openvoice/attentions.py:346: TracerWarning: Converting a tensor to a Python boolean might cause the trace to be incorrect. We can't record the data flow of Python values, so this value will be treated as a constant in the future. This means that the trace might not generalize to other inputs!
  pad_length = max(length - (self.window_size + 1), 0)
/opt/home/k8sworker/ci-ai/cibuilds/jobs/ov-notebook/jobs/OVNotebookOps/builds/835/archive/.workspace/scm/ov-notebook/notebooks/openvoice/OpenVoice/openvoice/attentions.py:347: TracerWarning: Converting a tensor to a Python boolean might cause the trace to be incorrect. We can't record the data flow of Python values, so this value will be treated as a constant in the future. This means that the trace might not generalize to other inputs!
  slice_start_position = max((self.window_size + 1) - length, 0)
/opt/home/k8sworker/ci-ai/cibuilds/jobs/ov-notebook/jobs/OVNotebookOps/builds/835/archive/.workspace/scm/ov-notebook/notebooks/openvoice/OpenVoice/openvoice/attentions.py:349: TracerWarning: Converting a tensor to a Python boolean might cause the trace to be incorrect. We can't record the data flow of Python values, so this value will be treated as a constant in the future. This means that the trace might not generalize to other inputs!
  if pad_length > 0:
/opt/home/k8sworker/ci-ai/cibuilds/jobs/ov-notebook/jobs/OVNotebookOps/builds/835/archive/.workspace/scm/ov-notebook/notebooks/openvoice/OpenVoice/openvoice/transforms.py:114: TracerWarning: Converting a tensor to a Python boolean might cause the trace to be incorrect. We can't record the data flow of Python values, so this value will be treated as a constant in the future. This means that the trace might not generalize to other inputs!
  if torch.min(inputs) < left or torch.max(inputs) > right:
/opt/home/k8sworker/ci-ai/cibuilds/jobs/ov-notebook/jobs/OVNotebookOps/builds/835/archive/.workspace/scm/ov-notebook/notebooks/openvoice/OpenVoice/openvoice/transforms.py:119: TracerWarning: Converting a tensor to a Python boolean might cause the trace to be incorrect. We can't record the data flow of Python values, so this value will be treated as a constant in the future. This means that the trace might not generalize to other inputs!
  if min_bin_width * num_bins > 1.0:
/opt/home/k8sworker/ci-ai/cibuilds/jobs/ov-notebook/jobs/OVNotebookOps/builds/835/archive/.workspace/scm/ov-notebook/notebooks/openvoice/OpenVoice/openvoice/transforms.py:121: TracerWarning: Converting a tensor to a Python boolean might cause the trace to be incorrect. We can't record the data flow of Python values, so this value will be treated as a constant in the future. This means that the trace might not generalize to other inputs!
  if min_bin_height * num_bins > 1.0:
/opt/home/k8sworker/ci-ai/cibuilds/jobs/ov-notebook/jobs/OVNotebookOps/builds/835/archive/.workspace/scm/ov-notebook/notebooks/openvoice/OpenVoice/openvoice/transforms.py:171: TracerWarning: Converting a tensor to a Python boolean might cause the trace to be incorrect. We can't record the data flow of Python values, so this value will be treated as a constant in the future. This means that the trace might not generalize to other inputs!
  assert (discriminant >= 0).all()
/opt/home/k8sworker/ci-ai/cibuilds/jobs/ov-notebook/jobs/OVNotebookOps/builds/835/archive/.workspace/scm/ov-notebook/.venv/lib/python3.8/site-packages/torch/jit/_trace.py:1303: TracerWarning: Trace had nondeterministic nodes. Did you forget call .eval() on your model? Nodes:
    %3293 : Float(1, 2, 43, strides=[86, 43, 1], requires_grad=0, device=cpu) = aten::randn(%3288, %3289, %3290, %3291, %3292) # /opt/home/k8sworker/ci-ai/cibuilds/jobs/ov-notebook/jobs/OVNotebookOps/builds/835/archive/.workspace/scm/ov-notebook/.venv/lib/python3.8/site-packages/nncf/torch/dynamic_graph/wrappers.py:86:0
    %5559 : Float(1, 192, 150, strides=[28800, 1, 192], requires_grad=0, device=cpu) = aten::randn_like(%m_p, %5554, %5555, %5556, %5557, %5558) # /opt/home/k8sworker/ci-ai/cibuilds/jobs/ov-notebook/jobs/OVNotebookOps/builds/835/archive/.workspace/scm/ov-notebook/.venv/lib/python3.8/site-packages/nncf/torch/dynamic_graph/wrappers.py:86:0
This may cause errors in trace checking. To disable trace checking, pass check_trace=False to torch.jit.trace()
  _check_trace(
/opt/home/k8sworker/ci-ai/cibuilds/jobs/ov-notebook/jobs/OVNotebookOps/builds/835/archive/.workspace/scm/ov-notebook/.venv/lib/python3.8/site-packages/torch/jit/_trace.py:1303: TracerWarning: Output nr 1. of the traced function does not match the corresponding output of the Python function. Detailed error:
Tensor-likes are not close!

Mismatched elements: 38094 / 39424 (96.6%)
Greatest absolute difference: 0.7026380896568298 at index (0, 0, 4174) (up to 1e-05 allowed)
Greatest relative difference: 43899.56701030928 at index (0, 0, 2529) (up to 1e-05 allowed)
  _check_trace(
/opt/home/k8sworker/ci-ai/cibuilds/jobs/ov-notebook/jobs/OVNotebookOps/builds/835/archive/.workspace/scm/ov-notebook/.venv/lib/python3.8/site-packages/torch/jit/_trace.py:1303: TracerWarning: Output nr 2. of the traced function does not match the corresponding output of the Python function. Detailed error:
Tensor-likes are not close!

Mismatched elements: 42 / 6622 (0.6%)
Greatest absolute difference: 1.0 at index (0, 0, 7, 1) (up to 1e-05 allowed)
Greatest relative difference: inf at index (0, 0, 7, 2) (up to 1e-05 allowed)
  _check_trace(

INFO:nncf:Statistics of the bitwidth distribution:
┍━━━━━━━━━━━━━━━━┯━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┯━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┑
│   Num bits (N) │ % all parameters (layers)   │ % ratio-defining parameters (layers)   │
┝━━━━━━━━━━━━━━━━┿━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┿━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┥
│              8 │ 100% (199 / 199)            │ 100% (199 / 199)                       │
┕━━━━━━━━━━━━━━━━┷━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┷━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┙

Output()

WARNING:tensorflow:Please fix your imports. Module tensorflow.python.training.tracking.base has been moved to tensorflow.python.trackable.base. The old module will be deleted in version 2.11.

/opt/home/k8sworker/ci-ai/cibuilds/jobs/ov-notebook/jobs/OVNotebookOps/builds/835/archive/.workspace/scm/ov-notebook/.venv/lib/python3.8/site-packages/torch/jit/_trace.py:1303: TracerWarning: Trace had nondeterministic nodes. Did you forget call .eval() on your model? Nodes:
    %1596 : Float(1, 192, 238, strides=[91392, 238, 1], requires_grad=0, device=cpu) = aten::randn_like(%m, %1591, %1592, %1593, %1594, %1595) # /opt/home/k8sworker/ci-ai/cibuilds/jobs/ov-notebook/jobs/OVNotebookOps/builds/835/archive/.workspace/scm/ov-notebook/.venv/lib/python3.8/site-packages/nncf/torch/dynamic_graph/wrappers.py:86:0
This may cause errors in trace checking. To disable trace checking, pass check_trace=False to torch.jit.trace()
  _check_trace(
/opt/home/k8sworker/ci-ai/cibuilds/jobs/ov-notebook/jobs/OVNotebookOps/builds/835/archive/.workspace/scm/ov-notebook/.venv/lib/python3.8/site-packages/torch/jit/_trace.py:1303: TracerWarning: Output nr 1. of the traced function does not match the corresponding output of the Python function. Detailed error:
Tensor-likes are not close!

Mismatched elements: 48656 / 60928 (79.9%)
Greatest absolute difference: 0.025416764430701733 at index (0, 0, 24537) (up to 1e-05 allowed)
Greatest relative difference: 12712.519438444924 at index (0, 0, 33972) (up to 1e-05 allowed)
  _check_trace(

INFO:nncf:Statistics of the bitwidth distribution:
┍━━━━━━━━━━━━━━━━┯━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┯━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┑
│   Num bits (N) │ % all parameters (layers)   │ % ratio-defining parameters (layers)   │
┝━━━━━━━━━━━━━━━━┿━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┿━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┥
│              8 │ 100% (194 / 194)            │ 100% (194 / 194)                       │
┕━━━━━━━━━━━━━━━━┷━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┷━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┙

Output()

Inference

Select inference device

core = ov.Core()

device = device_widget()
device

Dropdown(description='Device:', index=1, options=('CPU', 'AUTO'), value='AUTO')

Select reference tone

First of all, select the reference tone of voice to which the generated text will be converted: your can select from existing ones, record your own by selecting record_manually or upload you own file by load_manually

REFERENCE_VOICES_PATH = f"{repo_dir}/resources/"
reference_speakers = [
    *[path for path in os.listdir(REFERENCE_VOICES_PATH) if os.path.splitext(path)[-1] == ".mp3"],
    "record_manually",
    "load_manually",
]

ref_speaker = widgets.Dropdown(
    options=reference_speakers,
    value=reference_speakers[0],
    description="reference voice from which tone color will be copied",
    disabled=False,
)

ref_speaker

Dropdown(description='reference voice from which tone color will be copied', options=('demo_speaker2.mp3', 'de…

OUTPUT_DIR = Path("outputs/")
OUTPUT_DIR.mkdir(exist_ok=True)

ref_speaker_path = f"{REFERENCE_VOICES_PATH}/{ref_speaker.value}"
allowed_audio_types = ".mp4,.mp3,.wav,.wma,.aac,.m4a,.m4b,.webm"

if ref_speaker.value == "record_manually":
    ref_speaker_path = OUTPUT_DIR / "custom_example_sample.webm"
    from ipywebrtc import AudioRecorder, CameraStream

    camera = CameraStream(constraints={"audio": True, "video": False})
    recorder = AudioRecorder(stream=camera, filename=ref_speaker_path, autosave=True)
    display(recorder)
elif ref_speaker.value == "load_manually":
    upload_ref = widgets.FileUpload(
        accept=allowed_audio_types,
        multiple=False,
        description="Select audio with reference voice",
    )
    display(upload_ref)

Play the reference voice sample before cloning it’s tone to another speech

def save_audio(voice_source: widgets.FileUpload, out_path: str):
    with open(out_path, "wb") as output_file:
        assert len(voice_source.value) > 0, "Please select audio file"
        output_file.write(voice_source.value[0]["content"])


if ref_speaker.value == "load_manually":
    ref_speaker_path = f"{OUTPUT_DIR}/{upload_ref.value[0].name}"
    save_audio(upload_ref, ref_speaker_path)

Audio(ref_speaker_path)

torch_hub_local = Path("torch_hub_local/")
%env TORCH_HOME={str(torch_hub_local.absolute())}

env: TORCH_HOME=/opt/home/k8sworker/ci-ai/cibuilds/jobs/ov-notebook/jobs/OVNotebookOps/builds/835/archive/.workspace/scm/ov-notebook/notebooks/openvoice/torch_hub_local

Load speaker embeddings

# second step to fix a problem with silero downloading and installing
import os
import zipfile

url = "https://github.com/snakers4/silero-vad/zipball/v3.0"

torch_hub_dir = torch_hub_local / "hub"
torch.hub.set_dir(torch_hub_dir.as_posix())

zip_filename = "v3.0.zip"
output_path = torch_hub_dir / "v3.0"
if not (torch_hub_dir / zip_filename).exists():
    download_file(url, directory=torch_hub_dir, filename=zip_filename)
    zip_ref = zipfile.ZipFile((torch_hub_dir / zip_filename).as_posix(), "r")
    zip_ref.extractall(path=output_path.as_posix())
    zip_ref.close()

v3_dirs = [d for d in output_path.iterdir() if "snakers4-silero-vad" in d.as_posix()]
if len(v3_dirs) > 0 and not (torch_hub_dir / "snakers4_silero-vad_v3.0").exists():
    v3_dir = str(v3_dirs[0])
    os.rename(str(v3_dirs[0]), (torch_hub_dir / "snakers4_silero-vad_v3.0").as_posix())

v3.0.zip: 0.00B [00:00, ?B/s]

en_source_default_se = torch.load(f"{en_suffix}/en_default_se.pth")
en_source_style_se = torch.load(f"{en_suffix}/en_style_se.pth")
zh_source_se = torch.load(f"{zh_suffix}/zh_default_se.pth") if enable_chinese_lang else None

target_se, audio_name = se_extractor.get_se(ref_speaker_path, tone_color_converter, target_dir=OUTPUT_DIR, vad=True)

/tmp/ipykernel_2212381/4049537152.py:1: FutureWarning: You are using torch.load with weights_only=False (the current default value), which uses the default pickle module implicitly. It is possible to construct malicious pickle data which will execute arbitrary code during unpickling (See pytorch/pytorch for more details). In a future release, the default value for weights_only will be flipped to True. This limits the functions that could be executed during unpickling. Arbitrary objects will no longer be allowed to be loaded via this mode unless they are explicitly allowlisted by the user via torch.serialization.add_safe_globals. We recommend you start setting weights_only=True for any use case where you don't have full control of the loaded file. Please open an issue on GitHub for any issues related to this experimental feature.
  en_source_default_se = torch.load(f"{en_suffix}/en_default_se.pth")
/tmp/ipykernel_2212381/4049537152.py:2: FutureWarning: You are using torch.load with weights_only=False (the current default value), which uses the default pickle module implicitly. It is possible to construct malicious pickle data which will execute arbitrary code during unpickling (See pytorch/pytorch for more details). In a future release, the default value for weights_only will be flipped to True. This limits the functions that could be executed during unpickling. Arbitrary objects will no longer be allowed to be loaded via this mode unless they are explicitly allowlisted by the user via torch.serialization.add_safe_globals. We recommend you start setting weights_only=True for any use case where you don't have full control of the loaded file. Please open an issue on GitHub for any issues related to this experimental feature.
  en_source_style_se = torch.load(f"{en_suffix}/en_style_se.pth")

OpenVoice version: v1

/opt/home/k8sworker/ci-ai/cibuilds/jobs/ov-notebook/jobs/OVNotebookOps/builds/835/archive/.workspace/scm/ov-notebook/.venv/lib/python3.8/site-packages/torch/nn/modules/module.py:1562: UserWarning: A window was not provided. A rectangular window will be applied,which is known to cause spectral leakage. Other windows such as torch.hann_window or torch.hamming_window can are recommended to reduce spectral leakage.To suppress this warning and use a rectangular window, explicitly set window=torch.ones(n_fft, device=<device>). (Triggered internally at ../aten/src/ATen/native/SpectralOps.cpp:836.)
  return forward_call(*args, **kwargs)

[(0.0, 8.21), (9.292, 13.106), (13.228, 16.466), (16.684, 29.49225)]
after vad: dur = 28.07

/opt/home/k8sworker/ci-ai/cibuilds/jobs/ov-notebook/jobs/OVNotebookOps/builds/835/archive/.workspace/scm/ov-notebook/.venv/lib/python3.8/site-packages/torch/functional.py:666: UserWarning: stft with return_complex=False is deprecated. In a future pytorch release, stft will return complex tensors for all inputs, and return_complex=False will raise an error.
Note: you can still call torch.view_as_real on the complex output to recover the old return format. (Triggered internally at ../aten/src/ATen/native/SpectralOps.cpp:873.)
  return _VF.stft(input, n_fft, hop_length, win_length, window,  # type: ignore[attr-defined]

Replace original infer methods of OpenVoiceBaseClass with optimized OpenVINO inference.

There are pre and post processings that are not traceable and could not be offloaded to OpenVINO, instead of writing such processing ourselves we will rely on the already existing ones. We just replace infer and voice conversion functions of OpenVoiceBaseClass so that the the most computationally expensive part is done in OpenVINO.

def get_pathched_infer(ov_model: ov.Model, device: str) -> callable:
    compiled_model = core.compile_model(ov_model, device)

    def infer_impl(x, x_lengths, sid, noise_scale, length_scale, noise_scale_w):
        ov_output = compiled_model((x, x_lengths, sid, noise_scale, length_scale, noise_scale_w))
        return (torch.tensor(ov_output[0]),)

    return infer_impl


def get_patched_voice_conversion(ov_model: ov.Model, device: str) -> callable:
    compiled_model = core.compile_model(ov_model, device)

    def voice_conversion_impl(y, y_lengths, sid_src, sid_tgt, tau):
        ov_output = compiled_model((y, y_lengths, sid_src, sid_tgt, tau))
        return (torch.tensor(ov_output[0]),)

    return voice_conversion_impl


en_base_speaker_tts.model.infer = get_pathched_infer(ov_en_tts, device.value)
tone_color_converter.model.voice_conversion = get_patched_voice_conversion(ov_voice_conversion, device.value)
if enable_chinese_lang:
    zh_base_speaker_tts.model.infer = get_pathched_infer(ov_zh_tts, device.value)

Run inference

voice_source = widgets.Dropdown(
    options=["use TTS", "choose_manually"],
    value="use TTS",
    description="Voice source",
    disabled=False,
)

voice_source

Dropdown(description='Voice source', options=('use TTS', 'choose_manually'), value='use TTS')

if voice_source.value == "choose_manually":
    upload_orig_voice = widgets.FileUpload(
        accept=allowed_audio_types,
        multiple=False,
        description="audo whose tone will be replaced",
    )
    display(upload_orig_voice)

if voice_source.value == "choose_manually":
    orig_voice_path = f"{OUTPUT_DIR}/{upload_orig_voice.value[0].name}"
    save_audio(upload_orig_voice, orig_voice_path)
    source_se, _ = se_extractor.get_se(orig_voice_path, tone_color_converter, target_dir=OUTPUT_DIR, vad=True)
else:
    text = """
    OpenVINO toolkit is a comprehensive toolkit for quickly developing applications and solutions that solve
    a variety of tasks including emulation of human vision, automatic speech recognition, natural language processing,
    recommendation systems, and many others.
    """
    source_se = en_source_default_se
    orig_voice_path = OUTPUT_DIR / "tmp.wav"
    en_base_speaker_tts.tts(text, orig_voice_path, speaker="default", language="English")

 > Text splitted to sentences.
OpenVINO toolkit is a comprehensive toolkit for quickly developing applications and solutions that solve a variety of tasks including emulation of human vision,
automatic speech recognition, natural language processing, recommendation systems, and many others.
 > ===========================
ˈoʊpən vino* toolkit* ɪz ə ˌkɑmpɹiˈhɛnsɪv toolkit* fəɹ kˈwɪkli dɪˈvɛləpɪŋ ˌæpləˈkeɪʃənz ənd səˈluʃənz ðət sɑɫv ə vəɹˈaɪəti əv tæsks ˌɪnˈkludɪŋ ˌɛmjəˈleɪʃən əv ˈjumən ˈvɪʒən,
 length:173
 length:173
ˌɔtəˈmætɪk spitʃ ˌɹɛkɪgˈnɪʃən, ˈnætʃəɹəɫ ˈlæŋgwɪdʒ ˈpɹɑsɛsɪŋ, ˌɹɛkəmənˈdeɪʃən ˈsɪstəmz, ənd ˈmɛni ˈəðəɹz.
 length:105
 length:105

And finally, run voice tone conversion with OpenVINO optimized model

tau_slider = widgets.FloatSlider(
    value=0.3,
    min=0.01,
    max=2.0,
    step=0.01,
    description="tau",
    disabled=False,
    readout_format=".2f",
)
tau_slider

FloatSlider(value=0.3, description='tau', max=2.0, min=0.01, step=0.01)

resulting_voice_path = OUTPUT_DIR / "output_with_cloned_voice_tone.wav"

tone_color_converter.convert(
    audio_src_path=orig_voice_path,
    src_se=source_se,
    tgt_se=target_se,
    output_path=resulting_voice_path,
    tau=tau_slider.value,
    message="@MyShell",
)

Audio(orig_voice_path)

Audio(resulting_voice_path)

Run OpenVoice Gradio interactive demo

We can also use Gradio app to run TTS and voice tone conversion online.

import gradio as gr
import langid

supported_languages = ["zh", "en"]


def predict_impl(
    prompt,
    style,
    audio_file_pth,
    agree,
    output_dir,
    tone_color_converter,
    en_tts_model,
    zh_tts_model,
    en_source_default_se,
    en_source_style_se,
    zh_source_se,
):
    text_hint = ""
    if not agree:
        text_hint += "[ERROR] Please accept the Terms & Condition!\n"
        gr.Warning("Please accept the Terms & Condition!")
        return (
            text_hint,
            None,
            None,
        )

    language_predicted = langid.classify(prompt)[0].strip()
    print(f"Detected language:{language_predicted}")

    if language_predicted not in supported_languages:
        text_hint += f"[ERROR] The detected language {language_predicted} for your input text is not in our Supported Languages: {supported_languages}\n"
        gr.Warning(f"The detected language {language_predicted} for your input text is not in our Supported Languages: {supported_languages}")

        return (
            text_hint,
            None,
        )

    if language_predicted == "zh":
        tts_model = zh_tts_model
        if zh_tts_model is None:
            gr.Warning("TTS model for Chinece language was not loaded please set 'enable_chinese_lang=True`")
            return (
                text_hint,
                None,
            )
        source_se = zh_source_se
        language = "Chinese"
        if style not in ["default"]:
            text_hint += f"[ERROR] The style {style} is not supported for Chinese, which should be in ['default']\n"
            gr.Warning(f"The style {style} is not supported for Chinese, which should be in ['default']")
            return (
                text_hint,
                None,
            )
    else:
        tts_model = en_tts_model
        if style == "default":
            source_se = en_source_default_se
        else:
            source_se = en_source_style_se
        language = "English"
        supported_styles = [
            "default",
            "whispering",
            "shouting",
            "excited",
            "cheerful",
            "terrified",
            "angry",
            "sad",
            "friendly",
        ]
        if style not in supported_styles:
            text_hint += f"[ERROR] The style {style} is not supported for English, which should be in {*supported_styles,}\n"
            gr.Warning(f"The style {style} is not supported for English, which should be in {*supported_styles,}")
            return (
                text_hint,
                None,
            )

    speaker_wav = audio_file_pth

    if len(prompt) < 2:
        text_hint += "[ERROR] Please give a longer prompt text \n"
        gr.Warning("Please give a longer prompt text")
        return (
            text_hint,
            None,
        )
    if len(prompt) > 200:
        text_hint += (
            "[ERROR] Text length limited to 200 characters for this demo, please try shorter text. You can clone our open-source repo and try for your usage \n"
        )
        gr.Warning("Text length limited to 200 characters for this demo, please try shorter text. You can clone our open-source repo for your usage")
        return (
            text_hint,
            None,
        )

    # note diffusion_conditioning not used on hifigan (default mode), it will be empty but need to pass it to model.inference
    try:
        target_se, audio_name = se_extractor.get_se(speaker_wav, tone_color_converter, target_dir=OUTPUT_DIR, vad=True)
    except Exception as e:
        text_hint += f"[ERROR] Get target tone color error {str(e)} \n"
        gr.Warning("[ERROR] Get target tone color error {str(e)} \n")
        return (
            text_hint,
            None,
        )

    src_path = f"{output_dir}/tmp.wav"
    tts_model.tts(prompt, src_path, speaker=style, language=language)

    save_path = f"{output_dir}/output.wav"
    encode_message = "@MyShell"
    tone_color_converter.convert(
        audio_src_path=src_path,
        src_se=source_se,
        tgt_se=target_se,
        output_path=save_path,
        message=encode_message,
    )

    text_hint += "Get response successfully \n"

    return (
        text_hint,
        src_path,
        save_path,
    )

from functools import partial


predict = partial(
    predict_impl,
    output_dir=OUTPUT_DIR,
    tone_color_converter=tone_color_converter,
    en_tts_model=en_base_speaker_tts,
    zh_tts_model=zh_base_speaker_tts,
    en_source_default_se=en_source_default_se,
    en_source_style_se=en_source_style_se,
    zh_source_se=zh_source_se,
)

if not Path("gradio_helper.py").exists():
    r = requests.get(url="https://raw.githubusercontent.com/openvinotoolkit/openvino_notebooks/latest/notebooks/openvoice/gradio_helper.py")
    open("gradio_helper.py", "w").write(r.text)

from gradio_helper import make_demo

demo = make_demo(fn=predict)

try:
    demo.queue(max_size=2).launch(debug=False, height=1000)
except Exception:
    demo.queue(max_size=2).launch(share=True, debug=False, height=1000)
# if you are launching remotely, specify server_name and server_port
# demo.launch(server_name='your server name', server_port='server port in int')
# Read more in the docs: https://gradio.app/docs/

/opt/home/k8sworker/ci-ai/cibuilds/jobs/ov-notebook/jobs/OVNotebookOps/builds/835/archive/.workspace/scm/ov-notebook/.venv/lib/python3.8/site-packages/gradio/components/dropdown.py:100: UserWarning: The max_choices parameter is ignored when multiselect is False.
  warnings.warn(

Running on local URL:  http://127.0.0.1:7860

To create a public link, set share=True in launch().

# please uncomment and run this cell for stopping gradio interface
# demo.close()

Cleanup

# import shutil
# shutil.rmtree(CKPT_BASE_PATH)
# shutil.rmtree(IRS_PATH)
# shutil.rmtree(OUTPUT_DIR)