ESPnet Model Zoo

Utilities managing the pretrained models created by ESPnet. This function is inspired by the Asteroid pretrained model function.

• From version 0.1.0, the huggingface models can be also used: https://huggingface.co/models?filter=espnet
• Zenodo community: https://zenodo.org/communities/espnet/
• Registered models: table.csv

Install

pip install torch
pip install espnet_model_zoo

Python API for inference

model_name in the following section should be huggingface_id or one of the tags in the table.csv. Or you can directly provide zenodo URL (e.g., https://zenodo.org/record/xxxxxxx/files/hogehoge.zip?download=1).

ASR

import soundfile
from espnet2.bin.asr_inference import Speech2Text
speech2text = Speech2Text.from_pretrained(
    "model_name",
    # Decoding parameters are not included in the model file
    maxlenratio=0.0,
    minlenratio=0.0,
    beam_size=20,
    ctc_weight=0.3,
    lm_weight=0.5,
    penalty=0.0,
    nbest=1
)
# Confirm the sampling rate is equal to that of the training corpus.
# If not, you need to resample the audio data before inputting to speech2text
speech, rate = soundfile.read("speech.wav")
nbests = speech2text(speech)

text, *_ = nbests[0]
print(text)

TTS

import soundfile
from espnet2.bin.tts_inference import Text2Speech
text2speech = Text2Speech.from_pretrained("model_name")
speech = text2speech("foobar")["wav"]
soundfile.write("out.wav", speech.numpy(), text2speech.fs, "PCM_16")

Speech separation

import soundfile
from espnet2.bin.enh_inference import SeparateSpeech
separate_speech = SeparateSpeech.from_pretrained(
    "model_name",
    # for segment-wise process on long speech
    segment_size=2.4,
    hop_size=0.8,
    normalize_segment_scale=False,
    show_progressbar=True,
    ref_channel=None,
    normalize_output_wav=True,
)
# Confirm the sampling rate is equal to that of the training corpus.
# If not, you need to resample the audio data before inputting to speech2text
speech, rate = soundfile.read("long_speech.wav")
waves = separate_speech(speech[None, ...], fs=rate)

This API allows processing both short audio samples and long audio samples. For long audio samples, you can set the value of arguments segment_size, hop_size (optionally normalize_segment_scale and show_progressbar) to perform segment-wise speech enhancement/separation on the input speech. Note that the segment-wise processing is disabled by default.

import soundfile
from espnet_model_zoo.downloader import ModelDownloader
from espnet2.bin.asr_inference import Speech2Text
d = ModelDownloader()
speech2text = Speech2Text(
    **d.download_and_unpack("model_name"),
    # Decoding parameters are not included in the model file
    maxlenratio=0.0,
    minlenratio=0.0,
    beam_size=20,
    ctc_weight=0.3,
    lm_weight=0.5,
    penalty=0.0,
    nbest=1
)

import soundfile
from espnet_model_zoo.downloader import ModelDownloader
from espnet2.bin.tts_inference import Text2Speech
d = ModelDownloader()
text2speech = Text2Speech(**d.download_and_unpack("model_name"))

import soundfile
from espnet_model_zoo.downloader import ModelDownloader
from espnet2.bin.enh_inference import SeparateSpeech
d = ModelDownloader()
separate_speech = SeparateSpeech(
    **d.download_and_unpack("model_name"),
    # for segment-wise process on long speech
    segment_size=2.4,
    hop_size=0.8,
    normalize_segment_scale=False,
    show_progressbar=True,
    ref_channel=None,
    normalize_output_wav=True,
)

Instruction for ModelDownloader

from espnet_model_zoo.downloader import ModelDownloader
d = ModelDownloader("~/.cache/espnet")  # Specify cachedir
d = ModelDownloader()  # <module_dir> is used as cachedir by default

To obtain a model, you need to give a huggingface_idmodel` or a tag , which is listed in table.csv.

>>> d.download_and_unpack("kamo-naoyuki/mini_an4_asr_train_raw_bpe_valid.acc.best")
{"asr_train_config": <config path>, "asr_model_file": <model path>, ...}

You can specify the revision if it's huggingface_id giving with @:

>>> d.download_and_unpack("kamo-naoyuki/mini_an4_asr_train_raw_bpe_valid.acc.best@<revision>")
{"asr_train_config": <config path>, "asr_model_file": <model path>, ...}

Note that if the model already exists, you can skip downloading and unpacking.

You can also get a model with certain conditions.

d.download_and_unpack(task="asr", corpus="wsj")

If multiple models are found with the condition, the last model is selected. You can also specify the condition using "version" option.

d.download_and_unpack(task="asr", corpus="wsj", version=-1)  # Get the last model
d.download_and_unpack(task="asr", corpus="wsj", version=-2)  # Get previous model

You can also obtain it from the URL directly.

d.download_and_unpack("https://zenodo.org/record/...")

If you need to use a local model file using this API, you can also give it.

d.download_and_unpack("./some/where/model.zip")

In this case, the contents are also expanded in the cache directory, but the model is identified by the file path, so if you move the model to somewhere and unpack again, it's treated as another model, thus the contents are expanded again at another place.

Query model names

You can view the model names from our Zenodo community, https://zenodo.org/communities/espnet/, or using query(). All information are written in table.csv.

d.query("name")

You can also show them with specifying certain conditions.

d.query("name", task="asr")

Command line tools

• espnet_model_zoo_query

  # Query model name
  espnet_model_zoo_query task=asr corpus=wsj
  # Show all model name
  espnet_model_zoo_query
  # Query the other key
  espnet_model_zoo_query --key url task=asr corpus=wsj
  

• espnet_model_zoo_download

  espnet_model_zoo_download <model_name>  # Print the path of the downloaded file
  espnet_model_zoo_download --unpack true <model_name>   # Print the path of unpacked files
  

• espnet_model_zoo_upload

  export ACCESS_TOKEN=<access_token>
  espnet_zenodo_upload \
      --file <packed_model> \
      --title <title> \
      --description <description> \
      --creator_name <your-git-account>
  

Use pretrained model in ESPnet recipe

# e.g. ASR WSJ task
git clone https://github.com/espnet/espnet
pip install -e .
cd egs2/wsj/asr1
./run.sh --skip_data_prep false --skip_train true --download_model kamo-naoyuki/wsj

Register your model

Huggingface

1. Upload your model using huggingface API

   Coming soon...

2. Create a Pull Request to modify table.csv

   The models registered in this table.csv, the model are tested in the CI. Indeed, the model can be downloaded without modification table.csv.

3. (Administrator does) Increment the third version number of setup.py, e.g. 0.0.3 -> 0.0.4

4. (Administrator does) Release new version

Zenodo (Obsolete)

1. Upload your model to Zenodo

   You need to signup to Zenodo and create an access token to upload models. You can upload your own model by using espnet_model_zoo_upload command freely, but we normally upload a model using recipes.

2. Create a Pull Request to modify table.csv

   You need to append your record at the last line.

3. (Administrator does) Increment the third version number of setup.py, e.g. 0.0.3 -> 0.0.4

4. (Administrator does) Release new version