v1.0

third_party/Matcha-TTS/configs/logger/wandb.yaml

+# https://wandb.ai
+
+wandb:
+  _target_: lightning.pytorch.loggers.wandb.WandbLogger
+  # name: "" # name of the run (normally generated by wandb)
+  save_dir: "${paths.output_dir}"
+  offline: False
+  id: null # pass correct id to resume experiment!
+  anonymous: null # enable anonymous logging
+  project: "lightning-hydra-template"
+  log_model: False # upload lightning ckpts
+  prefix: "" # a string to put at the beginning of metric keys
+  # entity: "" # set to name of your wandb team
+  group: ""
+  tags: []
+  job_type: ""

third_party/Matcha-TTS/configs/model/cfm/default.yaml

+name: CFM
+solver: euler
+sigma_min: 1e-4

third_party/Matcha-TTS/configs/model/decoder/default.yaml

+channels: [256, 256]
+dropout: 0.05
+attention_head_dim: 64
+n_blocks: 1
+num_mid_blocks: 2
+num_heads: 2
+act_fn: snakebeta

third_party/Matcha-TTS/configs/model/encoder/default.yaml

+encoder_type: RoPE Encoder
+encoder_params:
+  n_feats: ${model.n_feats}
+  n_channels: 192
+  filter_channels: 768
+  filter_channels_dp: 256
+  n_heads: 2
+  n_layers: 6
+  kernel_size: 3
+  p_dropout: 0.1
+  spk_emb_dim: 64
+  n_spks: 1
+  prenet: true
+
+duration_predictor_params:
+  filter_channels_dp: ${model.encoder.encoder_params.filter_channels_dp}
+  kernel_size: 3
+  p_dropout: ${model.encoder.encoder_params.p_dropout}

third_party/Matcha-TTS/configs/model/matcha.yaml

+defaults:
+  - _self_
+  - encoder: default.yaml
+  - decoder: default.yaml
+  - cfm: default.yaml
+  - optimizer: adam.yaml
+
+_target_: matcha.models.matcha_tts.MatchaTTS
+n_vocab: 178
+n_spks: ${data.n_spks}
+spk_emb_dim: 64
+n_feats: 80
+data_statistics: ${data.data_statistics}
+out_size: null # Must be divisible by 4
+prior_loss: true
+use_precomputed_durations: ${data.load_durations}

third_party/Matcha-TTS/configs/model/optimizer/adam.yaml

+_target_: torch.optim.Adam
+_partial_: true
+lr: 1e-4
+weight_decay: 0.0

third_party/Matcha-TTS/configs/paths/default.yaml

+# path to root directory
+# this requires PROJECT_ROOT environment variable to exist
+# you can replace it with "." if you want the root to be the current working directory
+root_dir: ${oc.env:PROJECT_ROOT}
+
+# path to data directory
+data_dir: ${paths.root_dir}/data/
+
+# path to logging directory
+log_dir: ${paths.root_dir}/logs/
+
+# path to output directory, created dynamically by hydra
+# path generation pattern is specified in `configs/hydra/default.yaml`
+# use it to store all files generated during the run, like ckpts and metrics
+output_dir: ${hydra:runtime.output_dir}
+
+# path to working directory
+work_dir: ${hydra:runtime.cwd}

third_party/Matcha-TTS/configs/train.yaml

+# @package _global_
+
+# specify here default configuration
+# order of defaults determines the order in which configs override each other
+defaults:
+  - _self_
+  - data: ljspeech
+  - model: matcha
+  - callbacks: default
+  - logger: tensorboard # set logger here or use command line (e.g. `python train.py logger=tensorboard`)
+  - trainer: default
+  - paths: default
+  - extras: default
+  - hydra: default
+
+  # experiment configs allow for version control of specific hyperparameters
+  # e.g. best hyperparameters for given model and datamodule
+  - experiment: null
+
+  # config for hyperparameter optimization
+  - hparams_search: null
+
+  # optional local config for machine/user specific settings
+  # it's optional since it doesn't need to exist and is excluded from version control
+  - optional local: default
+
+  # debugging config (enable through command line, e.g. `python train.py debug=default)
+  - debug: null
+
+# task name, determines output directory path
+task_name: "train"
+
+run_name: ???
+
+# tags to help you identify your experiments
+# you can overwrite this in experiment configs
+# overwrite from command line with `python train.py tags="[first_tag, second_tag]"`
+tags: ["dev"]
+
+# set False to skip model training
+train: True
+
+# evaluate on test set, using best model weights achieved during training
+# lightning chooses best weights based on the metric specified in checkpoint callback
+test: True
+
+# simply provide checkpoint path to resume training
+ckpt_path: null
+
+# seed for random number generators in pytorch, numpy and python.random
+seed: 1234

third_party/Matcha-TTS/configs/trainer/cpu.yaml

+defaults:
+  - default
+
+accelerator: cpu
+devices: 1

third_party/Matcha-TTS/configs/trainer/ddp.yaml

+defaults:
+  - default
+
+strategy: ddp
+
+accelerator: gpu
+devices: [0,1]
+num_nodes: 1
+sync_batchnorm: True

third_party/Matcha-TTS/configs/trainer/ddp_sim.yaml

+defaults:
+  - default
+
+# simulate DDP on CPU, useful for debugging
+accelerator: cpu
+devices: 2
+strategy: ddp_spawn

third_party/Matcha-TTS/configs/trainer/default.yaml

+_target_: lightning.pytorch.trainer.Trainer
+
+default_root_dir: ${paths.output_dir}
+
+max_epochs: -1
+
+accelerator: gpu
+devices: [0]
+
+# mixed precision for extra speed-up
+precision: 16-mixed
+
+# perform a validation loop every N training epochs
+check_val_every_n_epoch: 1
+
+# set True to to ensure deterministic results
+# makes training slower but gives more reproducibility than just setting seeds
+deterministic: False
+
+gradient_clip_val: 5.0

third_party/Matcha-TTS/configs/trainer/gpu.yaml

+defaults:
+  - default
+
+accelerator: gpu
+devices: 1

third_party/Matcha-TTS/configs/trainer/mps.yaml

+defaults:
+  - default
+
+accelerator: mps
+devices: 1

third_party/Matcha-TTS/matcha/VERSION

+0.0.7.2

third_party/Matcha-TTS/matcha/app.py

+import tempfile
+from argparse import Namespace
+from pathlib import Path
+
+import gradio as gr
+import soundfile as sf
+import torch
+
+from matcha.cli import (
+    MATCHA_URLS,
+    VOCODER_URLS,
+    assert_model_downloaded,
+    get_device,
+    load_matcha,
+    load_vocoder,
+    process_text,
+    to_waveform,
+)
+from matcha.utils.utils import get_user_data_dir, plot_tensor
+
+LOCATION = Path(get_user_data_dir())
+
+args = Namespace(
+    cpu=False,
+    model="matcha_vctk",
+    vocoder="hifigan_univ_v1",
+    spk=0,
+)
+
+CURRENTLY_LOADED_MODEL = args.model
+
+
+def MATCHA_TTS_LOC(x):
+    return LOCATION / f"{x}.ckpt"
+
+
+def VOCODER_LOC(x):
+    return LOCATION / f"{x}"
+
+
+LOGO_URL = "https://shivammehta25.github.io/Matcha-TTS/images/logo.png"
+RADIO_OPTIONS = {
+    "Multi Speaker (VCTK)": {
+        "model": "matcha_vctk",
+        "vocoder": "hifigan_univ_v1",
+    },
+    "Single Speaker (LJ Speech)": {
+        "model": "matcha_ljspeech",
+        "vocoder": "hifigan_T2_v1",
+    },
+}
+
+# Ensure all the required models are downloaded
+assert_model_downloaded(MATCHA_TTS_LOC("matcha_ljspeech"), MATCHA_URLS["matcha_ljspeech"])
+assert_model_downloaded(VOCODER_LOC("hifigan_T2_v1"), VOCODER_URLS["hifigan_T2_v1"])
+assert_model_downloaded(MATCHA_TTS_LOC("matcha_vctk"), MATCHA_URLS["matcha_vctk"])
+assert_model_downloaded(VOCODER_LOC("hifigan_univ_v1"), VOCODER_URLS["hifigan_univ_v1"])
+
+device = get_device(args)
+
+# Load default model
+model = load_matcha(args.model, MATCHA_TTS_LOC(args.model), device)
+vocoder, denoiser = load_vocoder(args.vocoder, VOCODER_LOC(args.vocoder), device)
+
+
+def load_model(model_name, vocoder_name):
+    model = load_matcha(model_name, MATCHA_TTS_LOC(model_name), device)
+    vocoder, denoiser = load_vocoder(vocoder_name, VOCODER_LOC(vocoder_name), device)
+    return model, vocoder, denoiser
+
+
+def load_model_ui(model_type, textbox):
+    model_name, vocoder_name = RADIO_OPTIONS[model_type]["model"], RADIO_OPTIONS[model_type]["vocoder"]
+
+    global model, vocoder, denoiser, CURRENTLY_LOADED_MODEL  # pylint: disable=global-statement
+    if CURRENTLY_LOADED_MODEL != model_name:
+        model, vocoder, denoiser = load_model(model_name, vocoder_name)
+        CURRENTLY_LOADED_MODEL = model_name
+
+    if model_name == "matcha_ljspeech":
+        spk_slider = gr.update(visible=False, value=-1)
+        single_speaker_examples = gr.update(visible=True)
+        multi_speaker_examples = gr.update(visible=False)
+        length_scale = gr.update(value=0.95)
+    else:
+        spk_slider = gr.update(visible=True, value=0)
+        single_speaker_examples = gr.update(visible=False)
+        multi_speaker_examples = gr.update(visible=True)
+        length_scale = gr.update(value=0.85)
+
+    return (
+        textbox,
+        gr.update(interactive=True),
+        spk_slider,
+        single_speaker_examples,
+        multi_speaker_examples,
+        length_scale,
+    )
+
+
+@torch.inference_mode()
+def process_text_gradio(text):
+    output = process_text(1, text, device)
+    return output["x_phones"][1::2], output["x"], output["x_lengths"]
+
+
+@torch.inference_mode()
+def synthesise_mel(text, text_length, n_timesteps, temperature, length_scale, spk):
+    spk = torch.tensor([spk], device=device, dtype=torch.long) if spk >= 0 else None
+    output = model.synthesise(
+        text,
+        text_length,
+        n_timesteps=n_timesteps,
+        temperature=temperature,
+        spks=spk,
+        length_scale=length_scale,
+    )
+    output["waveform"] = to_waveform(output["mel"], vocoder, denoiser)
+    with tempfile.NamedTemporaryFile(suffix=".wav", delete=False) as fp:
+        sf.write(fp.name, output["waveform"], 22050, "PCM_24")
+
+    return fp.name, plot_tensor(output["mel"].squeeze().cpu().numpy())
+
+
+def multispeaker_example_cacher(text, n_timesteps, mel_temp, length_scale, spk):
+    global CURRENTLY_LOADED_MODEL  # pylint: disable=global-statement
+    if CURRENTLY_LOADED_MODEL != "matcha_vctk":
+        global model, vocoder, denoiser  # pylint: disable=global-statement
+        model, vocoder, denoiser = load_model("matcha_vctk", "hifigan_univ_v1")
+        CURRENTLY_LOADED_MODEL = "matcha_vctk"
+
+    phones, text, text_lengths = process_text_gradio(text)
+    audio, mel_spectrogram = synthesise_mel(text, text_lengths, n_timesteps, mel_temp, length_scale, spk)
+    return phones, audio, mel_spectrogram
+
+
+def ljspeech_example_cacher(text, n_timesteps, mel_temp, length_scale, spk=-1):
+    global CURRENTLY_LOADED_MODEL  # pylint: disable=global-statement
+    if CURRENTLY_LOADED_MODEL != "matcha_ljspeech":
+        global model, vocoder, denoiser  # pylint: disable=global-statement
+        model, vocoder, denoiser = load_model("matcha_ljspeech", "hifigan_T2_v1")
+        CURRENTLY_LOADED_MODEL = "matcha_ljspeech"
+
+    phones, text, text_lengths = process_text_gradio(text)
+    audio, mel_spectrogram = synthesise_mel(text, text_lengths, n_timesteps, mel_temp, length_scale, spk)
+    return phones, audio, mel_spectrogram
+
+
+def main():
+    description = """# 🍵 Matcha-TTS: A fast TTS architecture with conditional flow matching
+    ### [Shivam Mehta](https://www.kth.se/profile/smehta), [Ruibo Tu](https://www.kth.se/profile/ruibo), [Jonas Beskow](https://www.kth.se/profile/beskow), [Éva Székely](https://www.kth.se/profile/szekely), and [Gustav Eje Henter](https://people.kth.se/~ghe/)
+    We propose 🍵 Matcha-TTS, a new approach to non-autoregressive neural TTS, that uses conditional flow matching (similar to rectified flows) to speed up ODE-based speech synthesis. Our method:
+
+
+    * Is probabilistic
+    * Has compact memory footprint
+    * Sounds highly natural
+    * Is very fast to synthesise from
+
+
+    Check out our [demo page](https://shivammehta25.github.io/Matcha-TTS). Read our [arXiv preprint for more details](https://arxiv.org/abs/2309.03199).
+    Code is available in our [GitHub repository](https://github.com/shivammehta25/Matcha-TTS), along with pre-trained models.
+
+    Cached examples are available at the bottom of the page.
+    """
+
+    with gr.Blocks(title="🍵 Matcha-TTS: A fast TTS architecture with conditional flow matching") as demo:
+        processed_text = gr.State(value=None)
+        processed_text_len = gr.State(value=None)
+
+        with gr.Box():
+            with gr.Row():
+                gr.Markdown(description, scale=3)
+                with gr.Column():
+                    gr.Image(LOGO_URL, label="Matcha-TTS logo", height=50, width=50, scale=1, show_label=False)
+                    html = '<br><iframe width="560" height="315" src="https://www.youtube.com/embed/xmvJkz3bqw0?si=jN7ILyDsbPwJCGoa" title="YouTube video player" frameborder="0" allow="accelerometer; autoplay; clipboard-write; encrypted-media; gyroscope; picture-in-picture; web-share" allowfullscreen></iframe>'
+                    gr.HTML(html)
+
+        with gr.Box():
+            radio_options = list(RADIO_OPTIONS.keys())
+            model_type = gr.Radio(
+                radio_options, value=radio_options[0], label="Choose a Model", interactive=True, container=False
+            )
+
+            with gr.Row():
+                gr.Markdown("# Text Input")
+            with gr.Row():
+                text = gr.Textbox(value="", lines=2, label="Text to synthesise", scale=3)
+                spk_slider = gr.Slider(
+                    minimum=0, maximum=107, step=1, value=args.spk, label="Speaker ID", interactive=True, scale=1
+                )
+
+            with gr.Row():
+                gr.Markdown("### Hyper parameters")
+            with gr.Row():
+                n_timesteps = gr.Slider(
+                    label="Number of ODE steps",
+                    minimum=1,
+                    maximum=100,
+                    step=1,
+                    value=10,
+                    interactive=True,
+                )
+                length_scale = gr.Slider(
+                    label="Length scale (Speaking rate)",
+                    minimum=0.5,
+                    maximum=1.5,
+                    step=0.05,
+                    value=1.0,
+                    interactive=True,
+                )
+                mel_temp = gr.Slider(
+                    label="Sampling temperature",
+                    minimum=0.00,
+                    maximum=2.001,
+                    step=0.16675,
+                    value=0.667,
+                    interactive=True,
+                )
+
+                synth_btn = gr.Button("Synthesise")
+
+        with gr.Box():
+            with gr.Row():
+                gr.Markdown("### Phonetised text")
+                phonetised_text = gr.Textbox(interactive=False, scale=10, label="Phonetised text")
+
+        with gr.Box():
+            with gr.Row():
+                mel_spectrogram = gr.Image(interactive=False, label="mel spectrogram")
+
+                # with gr.Row():
+                audio = gr.Audio(interactive=False, label="Audio")
+
+        with gr.Row(visible=False) as example_row_lj_speech:
+            examples = gr.Examples(  # pylint: disable=unused-variable
+                examples=[
+                    [
+                        "We propose Matcha-TTS, a new approach to non-autoregressive neural TTS, that uses conditional flow matching (similar to rectified flows) to speed up O D E-based speech synthesis.",
+                        50,
+                        0.677,
+                        0.95,
+                    ],
+                    [
+                        "The Secret Service believed that it was very doubtful that any President would ride regularly in a vehicle with a fixed top, even though transparent.",
+                        2,
+                        0.677,
+                        0.95,
+                    ],
+                    [
+                        "The Secret Service believed that it was very doubtful that any President would ride regularly in a vehicle with a fixed top, even though transparent.",
+                        4,
+                        0.677,
+                        0.95,
+                    ],
+                    [
+                        "The Secret Service believed that it was very doubtful that any President would ride regularly in a vehicle with a fixed top, even though transparent.",
+                        10,
+                        0.677,
+                        0.95,
+                    ],
+                    [
+                        "The Secret Service believed that it was very doubtful that any President would ride regularly in a vehicle with a fixed top, even though transparent.",
+                        50,
+                        0.677,
+                        0.95,
+                    ],
+                    [
+                        "The narrative of these events is based largely on the recollections of the participants.",
+                        10,
+                        0.677,
+                        0.95,
+                    ],
+                    [
+                        "The jury did not believe him, and the verdict was for the defendants.",
+                        10,
+                        0.677,
+                        0.95,
+                    ],
+                ],
+                fn=ljspeech_example_cacher,
+                inputs=[text, n_timesteps, mel_temp, length_scale],
+                outputs=[phonetised_text, audio, mel_spectrogram],
+                cache_examples=True,
+            )
+
+        with gr.Row() as example_row_multispeaker:
+            multi_speaker_examples = gr.Examples(  # pylint: disable=unused-variable
+                examples=[
+                    [
+                        "Hello everyone! I am speaker 0 and I am here to tell you that Matcha-TTS is amazing!",
+                        10,
+                        0.677,
+                        0.85,
+                        0,
+                    ],
+                    [
+                        "Hello everyone! I am speaker 16 and I am here to tell you that Matcha-TTS is amazing!",
+                        10,
+                        0.677,
+                        0.85,
+                        16,
+                    ],
+                    [
+                        "Hello everyone! I am speaker 44 and I am here to tell you that Matcha-TTS is amazing!",
+                        50,
+                        0.677,
+                        0.85,
+                        44,
+                    ],
+                    [
+                        "Hello everyone! I am speaker 45 and I am here to tell you that Matcha-TTS is amazing!",
+                        50,
+                        0.677,
+                        0.85,
+                        45,
+                    ],
+                    [
+                        "Hello everyone! I am speaker 58 and I am here to tell you that Matcha-TTS is amazing!",
+                        4,
+                        0.677,
+                        0.85,
+                        58,
+                    ],
+                ],
+                fn=multispeaker_example_cacher,
+                inputs=[text, n_timesteps, mel_temp, length_scale, spk_slider],
+                outputs=[phonetised_text, audio, mel_spectrogram],
+                cache_examples=True,
+                label="Multi Speaker Examples",
+            )
+
+        model_type.change(lambda x: gr.update(interactive=False), inputs=[synth_btn], outputs=[synth_btn]).then(
+            load_model_ui,
+            inputs=[model_type, text],
+            outputs=[text, synth_btn, spk_slider, example_row_lj_speech, example_row_multispeaker, length_scale],
+        )
+
+        synth_btn.click(
+            fn=process_text_gradio,
+            inputs=[
+                text,
+            ],
+            outputs=[phonetised_text, processed_text, processed_text_len],
+            api_name="matcha_tts",
+            queue=True,
+        ).then(
+            fn=synthesise_mel,
+            inputs=[processed_text, processed_text_len, n_timesteps, mel_temp, length_scale, spk_slider],
+            outputs=[audio, mel_spectrogram],
+        )
+
+        demo.queue().launch(share=True)
+
+
+if __name__ == "__main__":
+    main()

third_party/Matcha-TTS/matcha/cli.py

+import argparse
+import datetime as dt
+import os
+import warnings
+from pathlib import Path
+
+import matplotlib.pyplot as plt
+import numpy as np
+import soundfile as sf
+import torch
+
+from matcha.hifigan.config import v1
+from matcha.hifigan.denoiser import Denoiser
+from matcha.hifigan.env import AttrDict
+from matcha.hifigan.models import Generator as HiFiGAN
+from matcha.models.matcha_tts import MatchaTTS
+from matcha.text import sequence_to_text, text_to_sequence
+from matcha.utils.utils import assert_model_downloaded, get_user_data_dir, intersperse
+
+MATCHA_URLS = {
+    "matcha_ljspeech": "https://github.com/shivammehta25/Matcha-TTS-checkpoints/releases/download/v1.0/matcha_ljspeech.ckpt",
+    "matcha_vctk": "https://github.com/shivammehta25/Matcha-TTS-checkpoints/releases/download/v1.0/matcha_vctk.ckpt",
+}
+
+VOCODER_URLS = {
+    "hifigan_T2_v1": "https://github.com/shivammehta25/Matcha-TTS-checkpoints/releases/download/v1.0/generator_v1",  # Old url: https://drive.google.com/file/d/14NENd4equCBLyyCSke114Mv6YR_j_uFs/view?usp=drive_link
+    "hifigan_univ_v1": "https://github.com/shivammehta25/Matcha-TTS-checkpoints/releases/download/v1.0/g_02500000",  # Old url: https://drive.google.com/file/d/1qpgI41wNXFcH-iKq1Y42JlBC9j0je8PW/view?usp=drive_link
+}
+
+MULTISPEAKER_MODEL = {
+    "matcha_vctk": {"vocoder": "hifigan_univ_v1", "speaking_rate": 0.85, "spk": 0, "spk_range": (0, 107)}
+}
+
+SINGLESPEAKER_MODEL = {"matcha_ljspeech": {"vocoder": "hifigan_T2_v1", "speaking_rate": 0.95, "spk": None}}
+
+
+def plot_spectrogram_to_numpy(spectrogram, filename):
+    fig, ax = plt.subplots(figsize=(12, 3))
+    im = ax.imshow(spectrogram, aspect="auto", origin="lower", interpolation="none")
+    plt.colorbar(im, ax=ax)
+    plt.xlabel("Frames")
+    plt.ylabel("Channels")
+    plt.title("Synthesised Mel-Spectrogram")
+    fig.canvas.draw()
+    plt.savefig(filename)
+
+
+def process_text(i: int, text: str, device: torch.device):
+    print(f"[{i}] - Input text: {text}")
+    x = torch.tensor(
+        intersperse(text_to_sequence(text, ["english_cleaners2"])[0], 0),
+        dtype=torch.long,
+        device=device,
+    )[None]
+    x_lengths = torch.tensor([x.shape[-1]], dtype=torch.long, device=device)
+    x_phones = sequence_to_text(x.squeeze(0).tolist())
+    print(f"[{i}] - Phonetised text: {x_phones[1::2]}")
+
+    return {"x_orig": text, "x": x, "x_lengths": x_lengths, "x_phones": x_phones}
+
+
+def get_texts(args):
+    if args.text:
+        texts = [args.text]
+    else:
+        with open(args.file, encoding="utf-8") as f:
+            texts = f.readlines()
+    return texts
+
+
+def assert_required_models_available(args):
+    save_dir = get_user_data_dir()
+    if not hasattr(args, "checkpoint_path") and args.checkpoint_path is None:
+        model_path = args.checkpoint_path
+    else:
+        model_path = save_dir / f"{args.model}.ckpt"
+        assert_model_downloaded(model_path, MATCHA_URLS[args.model])
+
+    vocoder_path = save_dir / f"{args.vocoder}"
+    assert_model_downloaded(vocoder_path, VOCODER_URLS[args.vocoder])
+    return {"matcha": model_path, "vocoder": vocoder_path}
+
+
+def load_hifigan(checkpoint_path, device):
+    h = AttrDict(v1)
+    hifigan = HiFiGAN(h).to(device)
+    hifigan.load_state_dict(torch.load(checkpoint_path, map_location=device)["generator"])
+    _ = hifigan.eval()
+    hifigan.remove_weight_norm()
+    return hifigan
+
+
+def load_vocoder(vocoder_name, checkpoint_path, device):
+    print(f"[!] Loading {vocoder_name}!")
+    vocoder = None
+    if vocoder_name in ("hifigan_T2_v1", "hifigan_univ_v1"):
+        vocoder = load_hifigan(checkpoint_path, device)
+    else:
+        raise NotImplementedError(
+            f"Vocoder {vocoder_name} not implemented! define a load_<<vocoder_name>> method for it"
+        )
+
+    denoiser = Denoiser(vocoder, mode="zeros")
+    print(f"[+] {vocoder_name} loaded!")
+    return vocoder, denoiser
+
+
+def load_matcha(model_name, checkpoint_path, device):
+    print(f"[!] Loading {model_name}!")
+    model = MatchaTTS.load_from_checkpoint(checkpoint_path, map_location=device)
+    _ = model.eval()
+
+    print(f"[+] {model_name} loaded!")
+    return model
+
+
+def to_waveform(mel, vocoder, denoiser=None, denoiser_strength=0.00025):
+    audio = vocoder(mel).clamp(-1, 1)
+    if denoiser is not None:
+        audio = denoiser(audio.squeeze(), strength=denoiser_strength).cpu().squeeze()
+
+    return audio.cpu().squeeze()
+
+
+def save_to_folder(filename: str, output: dict, folder: str):
+    folder = Path(folder)
+    folder.mkdir(exist_ok=True, parents=True)
+    plot_spectrogram_to_numpy(np.array(output["mel"].squeeze().float().cpu()), f"{filename}.png")
+    np.save(folder / f"{filename}", output["mel"].cpu().numpy())
+    sf.write(folder / f"{filename}.wav", output["waveform"], 22050, "PCM_24")
+    return folder.resolve() / f"{filename}.wav"
+
+
+def validate_args(args):
+    assert (
+        args.text or args.file
+    ), "Either text or file must be provided Matcha-T(ea)TTS need sometext to whisk the waveforms."
+    assert args.temperature >= 0, "Sampling temperature cannot be negative"
+    assert args.steps > 0, "Number of ODE steps must be greater than 0"
+
+    if args.checkpoint_path is None:
+        # When using pretrained models
+        if args.model in SINGLESPEAKER_MODEL:
+            args = validate_args_for_single_speaker_model(args)
+
+        if args.model in MULTISPEAKER_MODEL:
+            args = validate_args_for_multispeaker_model(args)
+    else:
+        # When using a custom model
+        if args.vocoder != "hifigan_univ_v1":
+            warn_ = "[-] Using custom model checkpoint! I would suggest passing --vocoder hifigan_univ_v1, unless the custom model is trained on LJ Speech."
+            warnings.warn(warn_, UserWarning)
+        if args.speaking_rate is None:
+            args.speaking_rate = 1.0
+
+    if args.batched:
+        assert args.batch_size > 0, "Batch size must be greater than 0"
+    assert args.speaking_rate > 0, "Speaking rate must be greater than 0"
+
+    return args
+
+
+def validate_args_for_multispeaker_model(args):
+    if args.vocoder is not None:
+        if args.vocoder != MULTISPEAKER_MODEL[args.model]["vocoder"]:
+            warn_ = f"[-] Using {args.model} model! I would suggest passing --vocoder {MULTISPEAKER_MODEL[args.model]['vocoder']}"
+            warnings.warn(warn_, UserWarning)
+    else:
+        args.vocoder = MULTISPEAKER_MODEL[args.model]["vocoder"]
+
+    if args.speaking_rate is None:
+        args.speaking_rate = MULTISPEAKER_MODEL[args.model]["speaking_rate"]
+
+    spk_range = MULTISPEAKER_MODEL[args.model]["spk_range"]
+    if args.spk is not None:
+        assert (
+            args.spk >= spk_range[0] and args.spk <= spk_range[-1]
+        ), f"Speaker ID must be between {spk_range} for this model."
+    else:
+        available_spk_id = MULTISPEAKER_MODEL[args.model]["spk"]
+        warn_ = f"[!] Speaker ID not provided! Using speaker ID {available_spk_id}"
+        warnings.warn(warn_, UserWarning)
+        args.spk = available_spk_id
+
+    return args
+
+
+def validate_args_for_single_speaker_model(args):
+    if args.vocoder is not None:
+        if args.vocoder != SINGLESPEAKER_MODEL[args.model]["vocoder"]:
+            warn_ = f"[-] Using {args.model} model! I would suggest passing --vocoder {SINGLESPEAKER_MODEL[args.model]['vocoder']}"
+            warnings.warn(warn_, UserWarning)
+    else:
+        args.vocoder = SINGLESPEAKER_MODEL[args.model]["vocoder"]
+
+    if args.speaking_rate is None:
+        args.speaking_rate = SINGLESPEAKER_MODEL[args.model]["speaking_rate"]
+
+    if args.spk != SINGLESPEAKER_MODEL[args.model]["spk"]:
+        warn_ = f"[-] Ignoring speaker id {args.spk} for {args.model}"
+        warnings.warn(warn_, UserWarning)
+        args.spk = SINGLESPEAKER_MODEL[args.model]["spk"]
+
+    return args
+
+
+@torch.inference_mode()
+def cli():
+    parser = argparse.ArgumentParser(
+        description=" 🍵 Matcha-TTS: A fast TTS architecture with conditional flow matching"
+    )
+    parser.add_argument(
+        "--model",
+        type=str,
+        default="matcha_ljspeech",
+        help="Model to use",
+        choices=MATCHA_URLS.keys(),
+    )
+
+    parser.add_argument(
+        "--checkpoint_path",
+        type=str,
+        default=None,
+        help="Path to the custom model checkpoint",
+    )
+
+    parser.add_argument(
+        "--vocoder",
+        type=str,
+        default=None,
+        help="Vocoder to use (default: will use the one suggested with the pretrained model))",
+        choices=VOCODER_URLS.keys(),
+    )
+    parser.add_argument("--text", type=str, default=None, help="Text to synthesize")
+    parser.add_argument("--file", type=str, default=None, help="Text file to synthesize")
+    parser.add_argument("--spk", type=int, default=None, help="Speaker ID")
+    parser.add_argument(
+        "--temperature",
+        type=float,
+        default=0.667,
+        help="Variance of the x0 noise (default: 0.667)",
+    )
+    parser.add_argument(
+        "--speaking_rate",
+        type=float,
+        default=None,
+        help="change the speaking rate, a higher value means slower speaking rate (default: 1.0)",
+    )
+    parser.add_argument("--steps", type=int, default=10, help="Number of ODE steps  (default: 10)")
+    parser.add_argument("--cpu", action="store_true", help="Use CPU for inference (default: use GPU if available)")
+    parser.add_argument(
+        "--denoiser_strength",
+        type=float,
+        default=0.00025,
+        help="Strength of the vocoder bias denoiser (default: 0.00025)",
+    )
+    parser.add_argument(
+        "--output_folder",
+        type=str,
+        default=os.getcwd(),
+        help="Output folder to save results (default: current dir)",
+    )
+    parser.add_argument("--batched", action="store_true", help="Batched inference (default: False)")
+    parser.add_argument(
+        "--batch_size", type=int, default=32, help="Batch size only useful when --batched (default: 32)"
+    )
+
+    args = parser.parse_args()
+
+    args = validate_args(args)
+    device = get_device(args)
+    print_config(args)
+    paths = assert_required_models_available(args)
+
+    if args.checkpoint_path is not None:
+        print(f"[🍵] Loading custom model from {args.checkpoint_path}")
+        paths["matcha"] = args.checkpoint_path
+        args.model = "custom_model"
+
+    model = load_matcha(args.model, paths["matcha"], device)
+    vocoder, denoiser = load_vocoder(args.vocoder, paths["vocoder"], device)
+
+    texts = get_texts(args)
+
+    spk = torch.tensor([args.spk], device=device, dtype=torch.long) if args.spk is not None else None
+    if len(texts) == 1 or not args.batched:
+        unbatched_synthesis(args, device, model, vocoder, denoiser, texts, spk)
+    else:
+        batched_synthesis(args, device, model, vocoder, denoiser, texts, spk)
+
+
+class BatchedSynthesisDataset(torch.utils.data.Dataset):
+    def __init__(self, processed_texts):
+        self.processed_texts = processed_texts
+
+    def __len__(self):
+        return len(self.processed_texts)
+
+    def __getitem__(self, idx):
+        return self.processed_texts[idx]
+
+
+def batched_collate_fn(batch):
+    x = []
+    x_lengths = []
+
+    for b in batch:
+        x.append(b["x"].squeeze(0))
+        x_lengths.append(b["x_lengths"])
+
+    x = torch.nn.utils.rnn.pad_sequence(x, batch_first=True)
+    x_lengths = torch.concat(x_lengths, dim=0)
+    return {"x": x, "x_lengths": x_lengths}
+
+
+def batched_synthesis(args, device, model, vocoder, denoiser, texts, spk):
+    total_rtf = []
+    total_rtf_w = []
+    processed_text = [process_text(i, text, "cpu") for i, text in enumerate(texts)]
+    dataloader = torch.utils.data.DataLoader(
+        BatchedSynthesisDataset(processed_text),
+        batch_size=args.batch_size,
+        collate_fn=batched_collate_fn,
+        num_workers=8,
+    )
+    for i, batch in enumerate(dataloader):
+        i = i + 1
+        start_t = dt.datetime.now()
+        b = batch["x"].shape[0]
+        output = model.synthesise(
+            batch["x"].to(device),
+            batch["x_lengths"].to(device),
+            n_timesteps=args.steps,
+            temperature=args.temperature,
+            spks=spk.expand(b) if spk is not None else spk,
+            length_scale=args.speaking_rate,
+        )
+
+        output["waveform"] = to_waveform(output["mel"], vocoder, denoiser, args.denoiser_strength)
+        t = (dt.datetime.now() - start_t).total_seconds()
+        rtf_w = t * 22050 / (output["waveform"].shape[-1])
+        print(f"[🍵-Batch: {i}] Matcha-TTS RTF: {output['rtf']:.4f}")
+        print(f"[🍵-Batch: {i}] Matcha-TTS + VOCODER RTF: {rtf_w:.4f}")
+        total_rtf.append(output["rtf"])
+        total_rtf_w.append(rtf_w)
+        for j in range(output["mel"].shape[0]):
+            base_name = f"utterance_{j:03d}_speaker_{args.spk:03d}" if args.spk is not None else f"utterance_{j:03d}"
+            length = output["mel_lengths"][j]
+            new_dict = {"mel": output["mel"][j][:, :length], "waveform": output["waveform"][j][: length * 256]}
+            location = save_to_folder(base_name, new_dict, args.output_folder)
+            print(f"[🍵-{j}] Waveform saved: {location}")
+
+    print("".join(["="] * 100))
+    print(f"[🍵] Average Matcha-TTS RTF: {np.mean(total_rtf):.4f} ± {np.std(total_rtf)}")
+    print(f"[🍵] Average Matcha-TTS + VOCODER RTF: {np.mean(total_rtf_w):.4f} ± {np.std(total_rtf_w)}")
+    print("[🍵] Enjoy the freshly whisked 🍵 Matcha-TTS!")
+
+
+def unbatched_synthesis(args, device, model, vocoder, denoiser, texts, spk):
+    total_rtf = []
+    total_rtf_w = []
+    for i, text in enumerate(texts):
+        i = i + 1
+        base_name = f"utterance_{i:03d}_speaker_{args.spk:03d}" if args.spk is not None else f"utterance_{i:03d}"
+
+        print("".join(["="] * 100))
+        text = text.strip()
+        text_processed = process_text(i, text, device)
+
+        print(f"[🍵] Whisking Matcha-T(ea)TS for: {i}")
+        start_t = dt.datetime.now()
+        output = model.synthesise(
+            text_processed["x"],
+            text_processed["x_lengths"],
+            n_timesteps=args.steps,
+            temperature=args.temperature,
+            spks=spk,
+            length_scale=args.speaking_rate,
+        )
+        output["waveform"] = to_waveform(output["mel"], vocoder, denoiser, args.denoiser_strength)
+        # RTF with HiFiGAN
+        t = (dt.datetime.now() - start_t).total_seconds()
+        rtf_w = t * 22050 / (output["waveform"].shape[-1])
+        print(f"[🍵-{i}] Matcha-TTS RTF: {output['rtf']:.4f}")
+        print(f"[🍵-{i}] Matcha-TTS + VOCODER RTF: {rtf_w:.4f}")
+        total_rtf.append(output["rtf"])
+        total_rtf_w.append(rtf_w)
+
+        location = save_to_folder(base_name, output, args.output_folder)
+        print(f"[+] Waveform saved: {location}")
+
+    print("".join(["="] * 100))
+    print(f"[🍵] Average Matcha-TTS RTF: {np.mean(total_rtf):.4f} ± {np.std(total_rtf)}")
+    print(f"[🍵] Average Matcha-TTS + VOCODER RTF: {np.mean(total_rtf_w):.4f} ± {np.std(total_rtf_w)}")
+    print("[🍵] Enjoy the freshly whisked 🍵 Matcha-TTS!")
+
+
+def print_config(args):
+    print("[!] Configurations: ")
+    print(f"\t- Model: {args.model}")
+    print(f"\t- Vocoder: {args.vocoder}")
+    print(f"\t- Temperature: {args.temperature}")
+    print(f"\t- Speaking rate: {args.speaking_rate}")
+    print(f"\t- Number of ODE steps: {args.steps}")
+    print(f"\t- Speaker: {args.spk}")
+
+
+def get_device(args):
+    if torch.cuda.is_available() and not args.cpu:
+        print("[+] GPU Available! Using GPU")
+        device = torch.device("cuda")
+    else:
+        print("[-] GPU not available or forced CPU run! Using CPU")
+        device = torch.device("cpu")
+    return device
+
+
+if __name__ == "__main__":
+    cli()