Python scripts to compute audio and language models from voxforge.org speech data and many sources. Models that can be built include:
- Kaldi nnet3 chain audio models
- KenLM language models in ARPA format
- sequitur g2p models
- wav2letter++ models
Important: Please note that these scripts form in no way a complete application ready for end-user consumption. However, if you are a developer interested in natural language processing you may find some of them useful. Contributions, patches and pull requests are very welcome.
At the time of this writing, the scripts here are focused on building the English and German VoxForge models. However, there is no reason why they couldn't be used to build other language models as well, feel free to contribute support for those.
- Zamia Speech
- Table of Contents
- Download
- Get Started with our Pre-Trained Models
- Get Started with a Demo STT Service Packaged in Docker
- Requirements
- Setup Notes
- Speech Corpora
- Text Corpora
- Language Model
- Submission Review and Transcription
- Lexica/Dictionaries
- Kaldi Models (recommended)
- wav2letter models
- Audiobook Segmentation and Transcription (Manual)
- Audiobook Segmentation and Transcription (kaldi)
- Training Voices for Zamia-TTS
- Model Distribution
- License
- Authors
Created by gh-md-toc
We have various models plus source code and binaries for the tools used to build these models available for download. Everything is free and open source.
All our model and data downloads can be found here: Downloads
Our pre-built ASR models can be downloaded here: ASR Models
- Kaldi ASR, English:
kaldi-generic-en-tdnn_fLarge nnet3-chain factorized TDNN model, trained on ~1200 hours of audio. Has decent background noise resistance and can also be used on phone recordings. Should provide the best accuracy but is a bit more resource intensive than the other models.kaldi-generic-en-tdnn_spLarge nnet3-chain model, trained on ~1200 hours of audio. Has decent background noise resistance and can also be used on phone recordings. Less accurate but also slightly less resource intensive than thetddn_fmodel.kaldi-generic-en-tdnn_250Same as the larger models but less resource intensive, suitable for use in embedded applications (e.g. a RaspberryPi 3).kaldi-generic-en-tri2b_chainGMM Model, trained on the same data as the above two models - meant for auto segmentation tasks.
- Kaldi ASR, German:
kaldi-generic-de-tdnn_fLarge nnet3-chain model, trained on ~400 hours of audio. Has decent background noise resistance and can also be used on phone recordings.kaldi-generic-de-tdnn_250Same as the large model but less resource intensive, suitable for use in embedded applications (e.g. a RaspberryPi 3).kaldi-generic-de-tri2b_chainGMM Model, trained on the same data as the above two models - meant for auto segmentation tasks.
- wav2letter++, German:
w2l-generic-deLarge model, trained on ~400 hours of audio. Has decent background noise resistance and can also be used on phone recordings.
NOTE: It is important to realize that these models can and should be adapted to your application domain. See Model Adaptation for details.
Our dictionaries can be downloaded here: Dictionaries
- IPA UTF-8, English:
dict-en.ipaBased on CMUDict with many additional entries generated via Sequitur G2P.
- IPA UTF-8, German:
dict-de.ipaCreated manually from scratch with many additional auto-reviewed entries extracted from Wiktionary.
Our pre-built G2P models can be downloaded here: G2P Models
- Sequitur, English:
sequitur-dict-en.ipaSequitur G2P model trained on our English IPA dictionary (UTF8).
- Sequitur, German:
sequitur-dict-de.ipaSequitur G2P model trained on our German IPA dictionary (UTF8).
Our pre-built ARPA language models can be downloaded here: Language Models
- KenLM, order 4, English, ARPA:
generic_en_lang_model_small
- KenLM, order 6, English, ARPA:
generic_en_lang_model_large
- KenLM, order 4, German, ARPA:
generic_de_lang_model_small
- KenLM, order 6, German, ARPA:
generic_de_lang_model_large
- Zamia-Speech where we host all our scripts and other sources used to build our models.
- py-kaldi-asr Python wrapper around Kaldi's nnet3-chain decoder complete with example scripts on how to use our models in your application.
- Binary AI Packages
- Raspbian APT Repo Binary packages in Debian format for Raspbian 9 (stretch, armhf, Raspberry Pi 2/3)
- Debian APT Repo Binary packages in Debian format for Debian 9 (stretch, amd64)
- CentOS YUM Repo Binary packages in RPM format for CentOS 7 (x86_64)
- Source AI Packages
- CentOS 7 Source packages in SRPM format for CentOS 7
Download a few sample wave files
$ wget http://goofy.zamia.org/zamia-speech/misc/demo_wavs.tgz
--2018-06-23 16:46:28-- http://goofy.zamia.org/zamia-speech/misc/demo_wavs.tgz
Resolving goofy.zamia.org (goofy.zamia.org)... 78.47.65.20
Connecting to goofy.zamia.org (goofy.zamia.org)|78.47.65.20|:80... connected.
HTTP request sent, awaiting response... 200 OK
Length: 619852 (605K) [application/x-gzip]
Saving to: ‘demo_wavs.tgz’
demo_wavs.tgz 100%[==========================================================>] 605.32K 2.01MB/s in 0.3s
2018-06-23 16:46:28 (2.01 MB/s) - ‘demo_wavs.tgz’ saved [619852/619852]unpack them:
$ tar xfvz demo_wavs.tgz
demo1.wav
demo2.wav
demo3.wav
demo4.wavdownload the demo program
$ wget http://goofy.zamia.org/zamia-speech/misc/kaldi_decode_wav.py
--2018-06-23 16:47:53-- http://goofy.zamia.org/zamia-speech/misc/kaldi_decode_wav.py
Resolving goofy.zamia.org (goofy.zamia.org)... 78.47.65.20
Connecting to goofy.zamia.org (goofy.zamia.org)|78.47.65.20|:80... connected.
HTTP request sent, awaiting response... 200 OK
Length: 2469 (2.4K) [text/plain]
Saving to: ‘kaldi_decode_wav.py’
kaldi_decode_wav.py 100%[==========================================================>] 2.41K --.-KB/s in 0s
2018-06-23 16:47:53 (311 MB/s) - ‘kaldi_decode_wav.py’ saved [2469/2469]now run kaldi automatic speech recognition on the demo wav files:
$ python kaldi_decode_wav.py -v demo?.wav
DEBUG:root:/opt/kaldi/model/kaldi-generic-en-tdnn_sp loading model...
DEBUG:root:/opt/kaldi/model/kaldi-generic-en-tdnn_sp loading model... done, took 1.473226s.
DEBUG:root:/opt/kaldi/model/kaldi-generic-en-tdnn_sp creating decoder...
DEBUG:root:/opt/kaldi/model/kaldi-generic-en-tdnn_sp creating decoder... done, took 0.143928s.
DEBUG:root:demo1.wav decoding took 0.37s, likelyhood: 1.863645
i cannot follow you she said
DEBUG:root:demo2.wav decoding took 0.54s, likelyhood: 1.572326
i should like to engage just for one whole life in that
DEBUG:root:demo3.wav decoding took 0.42s, likelyhood: 1.709773
philip knew that she was not an indian
DEBUG:root:demo4.wav decoding took 1.06s, likelyhood: 1.715135
he also contented that better confidence was established by carrying no weapons Determine the name of your pulseaudio mic source:
$ pactl list sources
Source #0
State: SUSPENDED
Name: alsa_input.usb-C-Media_Electronics_Inc._USB_PnP_Sound_Device-00.analog-mono
Description: CM108 Audio Controller Analog Mono
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^download and run demo:
$ wget 'http://goofy.zamia.org/zamia-speech/misc/kaldi_decode_live.py'
$ python kaldi_decode_live.py -s 'CM108'
Kaldi live demo V0.2
Loading model from /opt/kaldi/model/kaldi-generic-en-tdnn_250 ...
Please speak.
hallo computer
switch on the radio please
please switch on the light
what about the weather in stuttgart
how are you
thank you
good bye To start the STT service on your local machine, execute:
$ docker pull quay.io/mpuels/docker-py-kaldi-asr-and-model:kaldi-generic-en-tdnn_sp-r20180611
$ docker run --rm -p 127.0.0.1:8080:80/tcp quay.io/mpuels/docker-py-kaldi-asr-and-model:kaldi-generic-en-tdnn_sp-r20180611To transfer an audio file for transcription to the service, in a second terminal, execute:
$ git clone https://github.com/mpuels/docker-py-kaldi-asr-and-model.git
$ conda env create -f environment.yml
$ source activate py-kaldi-asr-client
$ ./asr_client.py asr.wav
INFO:root: 0.005s: 4000 frames ( 0.250s) decoded, status=200.
...
INFO:root:19.146s: 152000 frames ( 9.500s) decoded, status=200.
INFO:root:27.136s: 153003 frames ( 9.563s) decoded, status=200.
INFO:root:*****************************************************************
INFO:root:** wavfn : asr.wav
INFO:root:** hstr : speech recognition system requires training where individuals to exercise political system
INFO:root:** confidence : -0.578844
INFO:root:** decoding time : 27.14s
INFO:root:*****************************************************************The Docker image in the example above is the result of stacking 4 images on top of each other:
-
debian:8: https://hub.docker.com/_/debian/
Note: probably incomplete.
- Python 2.7 with nltk, numpy, ...
- KenLM
- kaldi
- wav2letter++
- py-nltools
- sox
- ffmpeg
Dependencies installation example for Debian:
apt-get install build-essential pkg-config python-pip python-dev python-setuptools python-wheel ffmpeg sox libatlas-base-dev
# Create a symbolic link because one of the pip packages expect atlas in this location:
ln -s /usr/include/x86_64-linux-gnu/atlas /usr/include/atlas
pip install numpy nltk cython
pip install py-kaldi-asr py-nltools
Just some rough notes on the environment needed to get these scripts to run. This is in no way a complete set of instructions, just some hints to get you started.
[speech]
vf_login = <your voxforge login>
speech_arc = /home/bofh/projects/ai/data/speech/arc
speech_corpora = /home/bofh/projects/ai/data/speech/corpora
kaldi_root = /apps/kaldi-cuda
; facebook's wav2letter++
w2l_env_activate = /home/bofh/projects/ai/w2l/bin/activate
w2l_train = /home/bofh/projects/ai/w2l/src/wav2letter/build/Train
w2l_decoder = /home/bofh/projects/ai/w2l/src/wav2letter/build/Decoder
wav16 = /home/bofh/projects/ai/data/speech/16kHz
noise_dir = /home/bofh/projects/ai/data/speech/corpora/noise
europarl_de = /home/bofh/projects/ai/data/corpora/de/europarl-v7.de-en.de
parole_de = /home/bofh/projects/ai/data/corpora/de/German Parole Corpus/DE_Parole/
europarl_en = /home/bofh/projects/ai/data/corpora/en/europarl-v7.de-en.en
cornell_movie_dialogs = /home/bofh/projects/ai/data/corpora/en/cornell_movie_dialogs_corpus
web_questions = /home/bofh/projects/ai/data/corpora/en/WebQuestions
yahoo_answers = /home/bofh/projects/ai/data/corpora/en/YahooAnswers
europarl_fr = /home/bofh/projects/ai/data/corpora/fr/europarl-v7.fr-en.fr
est_republicain = /home/bofh/projects/ai/data/corpora/fr/est_republicain.txt
wiktionary_de = /home/bofh/projects/ai/data/corpora/de/dewiktionary-20180320-pages-meta-current.xml
[tts]
host = localhost
port = 8300Some scripts expect al local tmp directory to be present, located in the same directory where all the scripts live, i.e.
mkdir tmpThe following list contains speech corpora supported by this script collection.
-
Forschergeist (German, 2 hours):
- Download all .tgz files into the directory
<~/.speechrc:speech_arc>/forschergeist - unpack them into the directory
<~/.speechrc:speech_corpora>/forschergeist
- Download all .tgz files into the directory
-
German Speechdata Package Version 2 (German, 148 hours):
- Unpack the archive such that the directories
dev,test, andtrainare direct subdirectories of<~/.speechrc:speech_arc>/gspv2. - Then run run the script
./import_gspv2.pyto convert the corpus to the VoxForge format. The resulting corpus will be written to<~/.speechrc:speech_corpora>/gspv2.
- Unpack the archive such that the directories
-
- Download the tarball
- unpack it into the directory
<~/.speechrc:speech_corpora>/(it will generate anoisesubdirectory there)
-
LibriSpeech ASR (English, 475 hours):
- Download the set of 360 hours "clean" speech tarball
- Unpack the archive such that the directory
LibriSpeechis a direct subdirectory of<~/.speechrc:speech_arc>. - Then run run the script
./import_librispeech.pyto convert the corpus to the VoxForge format. The resulting corpus will be written to<~/.speechrc:speech_corpora>/librispeech.
-
The LJ Speech Dataset (English, 24 hours):
- Download the tarball
- Unpack the archive such that the directory
LJSpeech-1.1is a direct subdirectory of<~/.speechrc:speech_arc>. - Then run run the script
import_ljspeech.pyto convert the corpus to the VoxForge format. The resulting corpus will be written to<~/.speechrc:speech_corpora>/lindajohnson-11.
-
Mozilla Common Voice German (German, 140 hours):
- Download
de.tar.gz - Unpack the archive such that the directory
cv_deis a direct subdirectory of<~/.speechrc:speech_arc>. - Then run run the script
./import_mozde.pyto convert the corpus to the VoxForge format. The resulting corpus will be written to<~/.speechrc:speech_corpora>/cv_de.
- Download
-
Mozilla Common Voice V1 (English, 252 hours):
- Download
cv_corpus_v1.tar.gz - Unpack the archive such that the directory
cv_corpus_v1is a direct subdirectory of<~/.speechrc:speech_arc>. - Then run run the script
./import_mozcv1.pyto convert the corpus to the VoxForge format. The resulting corpus will be written to<~/.speechrc:speech_corpora>/cv_corpus_v1.
- Download
-
- Download
de_DE.tgz,en_UK.tgz,en_US.tgz,fr_FR.tgz(Mirror) - Create a subdirectory
m_ailabsin<~/.speechrc:speech_arc> - Unpack the downloaded tarbals inside the
m_ailabssubdirectory - For French, create a directory
by_bookand movemaleandfemaledirectories in it as the archive does not follow exactly English and German structures - Then run run the script
./import_mailabs.pyto convert the corpus to the VoxForge format. The resulting corpus will be written to<~/.speechrc:speech_corpora>/m_ailabs_en,<~/.speechrc:speech_corpora>/m_ailabs_deand<~/.speechrc:speech_corpora>/m_ailabs_fr.
- Download
-
TED-LIUM Release 3 (English, 210 hours):
- Download
TEDLIUM_release-3.tgz - Unpack the archive such that the directory
TEDLIUM_release-3is a direct subdirectory of<~/.speechrc:speech_arc>. - Then run run the script
./import_tedlium3.pyto convert the corpus to the VoxForge format. The resulting corpus will be written to<~/.speechrc:speech_corpora>/tedlium3.
- Download
-
- Download all .tgz files into the directory
<~/.speechrc:speech_arc>/voxforge_en - unpack them into the directory
<~/.speechrc:speech_corpora>/voxforge_en
- Download all .tgz files into the directory
-
- Download all .tgz files into the directory
<~/.speechrc:speech_arc>/voxforge_de - unpack them into the directory
<~/.speechrc:speech_corpora>/voxforge_de
- Download all .tgz files into the directory
-
- Download all .tgz files into the directory
<~/.speechrc:speech_arc>/voxforge_fr - unpack them into the directory
<~/.speechrc:speech_corpora>/voxforge_fr
- Download all .tgz files into the directory
-
- Download all .tgz files into the directory
<~/.speechrc:speech_arc>/zamia_en - unpack them into the directory
<~/.speechrc:speech_corpora>/zamia_en
- Download all .tgz files into the directory
-
- Download all .tgz files into the directory
<~/.speechrc:speech_arc>/zamia_de - unpack them into the directory
<~/.speechrc:speech_corpora>/zamia_de
- Download all .tgz files into the directory
Technical note: For most corpora we have corrected transcripts in our databases which can be found
in data/src/speech/<corpus_name>/transcripts_*.csv. As these have been created by many hours of (semi-)
manual review they should be of higher quality than the original prompts so they will be used during
training of our ASR models.
Once you have downloaded and, if necessary, converted a corpus you need to run
./speech_audio_scan.py <corpus name>on it. This will add missing prompts to the CSV databases and convert audio files to 16kHz mono WAVE format.
To improve noise resistance it is possible to derive corpora from existing ones with noise added:
./speech_gen_noisy.py zamia_de
./speech_audio_scan.py zamia_de_noisy
cp data/src/speech/zamia_de/spk2gender data/src/speech/zamia_de_noisy/
cp data/src/speech/zamia_de/spk_test.txt data/src/speech/zamia_de_noisy/
./auto_review.py -a zamia_de_noisy
./apply_review.py -l de zamia_de_noisy review-result.csv This script will run recording through typical telephone codecs. Such a corpus can be used to train models that support 8kHz phone recordings:
./speech_gen_phone.py zamia_de
./speech_audio_scan.py zamia_de_phone
cp data/src/speech/zamia_de/spk2gender data/src/speech/zamia_de_phone/
cp data/src/speech/zamia_de/spk_test.txt data/src/speech/zamia_de_phone/
./auto_review.py -a zamia_de_phone
./apply_review.py -l de zamia_de_phone review-result.csv The following list contains text corpora that can be used to train language models with the scripts contained in this repository:
-
Europarl, specifically parallel corpus German-English and parallel corpus French-English:
- corresponding variable in
.speechrc:europarl_de,europarl_en,europarl_fr - sentences extraction: run
./speech_sentences.py europarl_de,./speech_sentences.py europarl_enand./speech_sentences.py europarl_fr
- corresponding variable in
-
Cornell Movie--Dialogs Corpus:
- corresponding variable in
.speechrc:cornell_movie_dialogs - sentences extraction: run
./speech_sentences.py cornell_movie_dialogs
- corresponding variable in
-
- corresponding variable in
.speechrc:parole_de - sentences extraction: train punkt tokenizer using
./speech_train_punkt_tokenizer.py, then run./speech_sentences.py parole_de
- corresponding variable in
-
WebQuestions:
web_questions- corresponding variable in
.speechrc:web_questions - sentences extraction: run
./speech_sentences.py web_questions
- corresponding variable in
-
Yahoo! Answers dataset:
yahoo_answers- corresponding variable in
.speechrc:yahoo_answers - sentences extraction: run
./speech_sentences.py yahoo_answers
- corresponding variable in
-
CNRTL Est Républicain Corpus, large corpus of news articles (4.3M headlines/paragraphs) available under a CC BY-NC-SA license. Download XML files and extract headlines and paragraphs to a text file with the following command:
xmllint --xpath '//*[local-name()="div"][@type="article"]//*[local-name()="p" or local-name()="head"]/text()' Annee*/*.xml | perl -pe 's/^ +//g ; s/^ (.+)/$1\n/g ; chomp' > est_republicain.txt- corresponding variable in
.speechrc:est_republicain - sentences extraction: run
./speech_sentences.py est_republicain
- corresponding variable in
Sentences can also be extracted from our speech corpora. To do that, run:
-
English Speech Corpora
./speech_sentences.py voxforge_en./speech_sentences.py librispeech./speech_sentences.py zamia_en./speech_sentences.py cv_corpus_v1./speech_sentences.py ljspeech./speech_sentences.py m_ailabs_en./speech_sentences.py tedlium3
-
German Speech Corpora
./speech_sentences.py forschergeist./speech_sentences.py gspv2./speech_sentences.py voxforge_de./speech_sentences.py zamia_de./speech_sentences.py m_ailabs_de./speech_sentences.py cv_de
Prerequisites:
- text corpora
europarl_en,cornell_movie_dialogs,web_questions, andyahoo_answersare installed, sentences extracted (see instructions above). - sentences are extracted from speech corpora
librispeech,voxforge_en,zamia_en,cv_corpus_v1,ljspeech,m_ailabs_en,tedlium3
To train a small, pruned English language model of order 4 using KenLM for use in both kaldi and wav2letter builds run:
./speech_build_lm.py generic_en_lang_model_small europarl_en cornell_movie_dialogs web_questions yahoo_answers librispeech voxforge_en zamia_en cv_corpus_v1 ljspeech m_ailabs_en tedlium3to train a larger model of order 6 with less pruning:
./speech_build_lm.py -o 6 -p "0 0 0 0 1" generic_en_lang_model_large europarl_en cornell_movie_dialogs web_questions yahoo_answers librispeech voxforge_en zamia_en cv_corpus_v1 ljspeech m_ailabs_en tedlium3to train a medium size model of order 5:
./speech_build_lm.py -o 5 -p "0 0 1 2" generic_en_lang_model_medium europarl_en cornell_movie_dialogs web_questions yahoo_answers librispeech voxforge_en zamia_en cv_corpus_v1 ljspeech m_ailabs_en tedlium3Prerequisites:
- text corpora
europarl_deandparole_deare installed, sentences extracted (see instructions above). - sentences are extracted from speech corpora
forschergeist,gspv2,voxforge_de,zamia_de,m_ailabs_de,cv_de
To train a small, pruned German language model of order 4 using KenLM for use in both kaldi and wav2letter builds run:
./speech_build_lm.py generic_de_lang_model_small europarl_de parole_de forschergeist gspv2 voxforge_de zamia_de m_ailabs_de cv_deto train a larger model of order 6 with less pruning:
./speech_build_lm.py -o 6 -p "0 0 0 0 1" generic_de_lang_model_large europarl_de parole_de forschergeist gspv2 voxforge_de zamia_de m_ailabs_de cv_deto train a medium size model of order 5:
./speech_build_lm.py -o 5 -p "0 0 1 2" generic_de_lang_model_medium europarl_de parole_de forschergeist gspv2 voxforge_de zamia_de m_ailabs_de cv_dePrerequisites:
- text corpora
europarl_frandest_republicainare installed, sentences extracted (see instructions above). - sentences are extracted from speech corpora
voxforge_frandm_ailabs_fr
To train a French language model using KenLM run:
./speech_build_lm.py generic_fr_lang_model europarl_fr est_republicain voxforge_fr m_ailabs_frThe main tool used for submission review, transcription and lexicon expansion is:
./speech_editor.pyNOTE: We use the terms lexicon and dictionary interchangably in this documentation and our scripts.
Currently, we have two lexica, one for English and one for German (in data/src/dicts):
-
dict-en.ipa
- English
- originally based on The CMU Pronouncing Dictionary (http://www.speech.cs.cmu.edu/cgi-bin/cmudict)
- additional manual and Sequitur G2P based entries
-
dict-de.ipa
- started manually from scratch
- once enough entries existed to train a reasonable Sequitur G2P model, many entries where converted from German wiktionary (see below)
The native format of our lexica is in (UTF8) IPA with semicolons as separator. This format is then converted to whatever format is used by the target ASR engine by the corresponding export scripts.
Many lexicon-related tools rely on Sequitur G2P to compute pronunciations for words missing from the dictionary. The
necessary models can be downloaded from our file server: http://goofy.zamia.org/zamia-speech/g2p/ .
For installation, download and unpack them and then put links to them under data/models like so:
data/models/sequitur-dict-de.ipa-latest -> <your model dir>/sequitur-dict-de.ipa-r20180510
data/models/sequitur-dict-en.ipa-latest -> <your model dir>/sequitur-dict-en.ipa-r20180510To train your own Sequitur G2P models, use the export and train scripts provided, e.g.:
[guenter@dagobert speech]$ ./speech_sequitur_export.py -d dict-de.ipa
INFO:root:loading lexicon...
INFO:root:loading lexicon...done.
INFO:root:sequitur workdir data/dst/dict-models/dict-de.ipa/sequitur done.
[guenter@dagobert speech]$ ./speech_sequitur_train.sh dict-de.ipa
training sample: 322760 + 16988 devel
iteration: 0
..../speech_lex_edit.py word [word2 ...]is the main curses based, interactive lexicon editor. It will automatically produce candidate entries for new words using Sequitur G2P, MaryTTS and eSpeakNG. The user can then edit these entries manually if necessary and check them by listening to them being synthesized via MaryTTS in different voices.
The lexicon editor is also integrated into various other tools, speech_editor.py in particular
which allows you to transcribe, review and add missing words for new audio samples
within one tool - which is recommended.
I also tend to review lexicon entries randomly from time to time. For that I have a small script which will pick 20 random entries where Sequitur G2P disagrees with the current transcription in the lexicon:
./speech_lex_edit.py `./speech_lex_review.py`Also, I sometimes use this command to add missing words from transcripts in batch mode:
./speech_lex_edit.py `./speech_lex_missing.py`For the German lexicon, entries can be extracted from the German wiktionary using a set of scripts. To do that, the first step is to extract a set of candidate entries from an wiktionary XML dump:
./wiktionary_extract_ipa.py this will output extracted entries to data/dst/speech/de/dict_wiktionary_de.txt. We now need to
train a Sequitur G2P model that translates these entries into our own IPA style and phoneme set:
./wiktionary_sequitur_export.py
./wiktionary_sequitur_train.shfinally, we translate the entries and check them against the predictions from our regular Sequitur G2P model:
./wiktionary_sequitur_gen.pythis script produces two output files: data/dst/speech/de/dict_wiktionary_gen.txt contains acceptable entries,
data/dst/speech/de/dict_wiktionary_rej.txt contains rejected entries.
The following recipe trains Kaldi models for English.
Before running it, make sure all prerequisites are met (see above for instructions on these):
- language model
generic_en_lang_model_smallbuilt - some or all speech corpora of
voxforge_en,librispeech,cv_corpus_v1,ljspeech,m_ailabs_en,tedlium3andzamia_enare installed, converted and scanned. - optionally noise augmented corpora:
voxforge_en_noisy,voxforge_en_phone,librispeech_en_noisy,librispeech_en_phone,cv_corpus_v1_noisy,cv_corpus_v1_phone,zamia_en_noisyandzamia_en_phone
./speech_kaldi_export.py generic-en-small dict-en.ipa generic_en_lang_model_small voxforge_en librispeech zamia_en
cd data/dst/asr-models/kaldi/generic-en-small
./run-chain.shexport run with noise augmented corpora included:
./speech_kaldi_export.py generic-en dict-en.ipa generic_en_lang_model_small voxforge_en cv_corpus_v1 librispeech ljspeech m_ailabs_en tedlium3 zamia_en voxforge_en_noisy librispeech_noisy cv_corpus_v1_noisy cv_corpus_v1_phone zamia_en_noisy voxforge_en_phone librispeech_phone zamia_en_phoneThe following recipe trains Kaldi models for German.
Before running it, make sure all prerequisites are met (see above for instructions on these):
- language model
generic_de_lang_model_smallbuilt - some or all speech corpora of
voxforge_de,gspv2,forschergeist,zamia_de,m_ailabs_de,cv_deare installed, converted and scanned. - optionally noise augmented corpora:
voxforge_de_noisy,voxforge_de_phone,zamia_de_noisyandzamia_de_phone
./speech_kaldi_export.py generic-de-small dict-de.ipa generic_de_lang_model_small voxforge_de gspv2 [ forschergeist zamia_de ...]
cd data/dst/asr-models/kaldi/generic-de-small
./run-chain.shexport run with noise augmented corpora included:
./speech_kaldi_export.py generic-de dict-de.ipa generic_de_lang_model_small voxforge_de gspv2 forschergeist zamia_de voxforge_de_noisy voxforge_de_phone zamia_de_noisy zamia_de_phone m_ailabs_de cv_deFor a standalone kaldi model adaptation tool that does not require a complete zamia-speech setup, see
Existing kaldi models (such as the ones we provide for download but also those you may train from scratch using our scripts) can be adapted to (typically domain specific) language models, JSGF grammars and grammar FSTs.
Here is an example how to adapt our English model to a simple command and control JSGF grammar. Please note that this is just a toy example - for real world usage you will probably want to add garbage phoneme loops to the grammar or produce a language model that has some noise resistance built in right away.
Here is the grammar we will use:
#JSGF V1.0;
grammar org.zamia.control;
public <control> = <wake> | <politeCommand> ;
<wake> = ( good morning | hello | ok | activate ) computer;
<politeCommand> = [ please | kindly | could you ] <command> [ please | thanks | thank you ];
<command> = <onOffCommand> | <muteCommand> | <volumeCommand> | <weatherCommand>;
<onOffCommand> = [ turn | switch ] [the] ( light | fan | music | radio ) (on | off) ;
<volumeCommand> = turn ( up | down ) the ( volume | music | radio ) ;
<muteCommand> = mute the ( music | radio ) ;
<weatherCommand> = (what's | what) is the ( temperature | weather ) ;
the next step is to set up a kaldi model adaptation experiment using this script:
./speech_kaldi_adapt.py data/models/kaldi-generic-en-tdnn_250-latest dict-en.ipa control.jsgf control-enhere, data/models/kaldi-generic-en-tdnn_250-latest is the model to be adapted, dict-en.ipa is the dictionary which
will be used by the new model, control.jsgf is the JSGF grammar we want the model to be adapted to (you could specify an
FST source file or a language model instead here) and control-en is the name of the new model that will be created.
To run the actual adaptation, change into the model directory and run the adaptation script there:
cd data/dst/asr-models/kaldi/control-en
./run-adaptation.sh finally, you can create a tarball from the newly created model:
cd ../../../../..
./speech_dist.sh control-en kaldi adapt./wav2letter_export.py -l en -v generic-en dict-en.ipa generic_en_lang_model_large voxforge_en cv_corpus_v1 librispeech ljspeech m_ailabs_en tedlium3 zamia_en voxforge_en_noisy librispeech_noisy cv_corpus_v1_noisy cv_corpus_v1_phone zamia_en_noisy voxforge_en_phone librispeech_phone zamia_en_phone
pushd data/dst/asr-models/wav2letter/generic-en/
bash run_train.sh./wav2letter_export.py -l de -v generic-de dict-de.ipa generic_de_lang_model_large voxforge_de gspv2 forschergeist zamia_de voxforge_de_noisy voxforge_de_phone zamia_de_noisy zamia_de_phone m_ailabs_de cv_de
pushd data/dst/asr-models/wav2letter/generic-de/
bash run_train.shcreate auto-review case:
./wav2letter_auto_review.py -l de w2l-generic-de-latest gspv2run it:
pushd tmp/w2letter_auto_review
bash run_auto_review.sh
popdapply the results:
./wav2letter_apply_review.pySome notes on how to segment and transcribe audiobooks or other audio sources (e.g. from librivox) using the abook scripts provided:
MP3
```bash
ffmpeg -i foo.mp3 foo.wav
```
MKV
mkvextract tracks foo.mkv 0:foo.ogg
opusdec foo.ogg foo.wavsox foo.wav -r 16000 -c 1 -b 16 foo_16m.wavThis tool will use silence detection to find good cut-points. You may want to adjust its settings to achieve a good balance of short-segments but few words split in half.
./abook-segment.py foo_16m.wavsettings:
[guenter@dagobert speech]$ ./abook-segment.py -h
Usage: abook-segment.py [options] foo.wav
Options:
-h, --help show this help message and exit
-s SILENCE_LEVEL, --silence-level=SILENCE_LEVEL
silence level (default: 2048 / 65536)
-l MIN_SIL_LENGTH, --min-sil-length=MIN_SIL_LENGTH
minimum silence length (default: 0.07s)
-m MIN_UTT_LENGTH, --min-utt-length=MIN_UTT_LENGTH
minimum utterance length (default: 2.00s)
-M MAX_UTT_LENGTH, --max-utt-length=MAX_UTT_LENGTH
maximum utterance length (default: 9.00s)
-o OUTDIRFN, --out-dir=OUTDIRFN
output directory (default: abook/segments)
-v, --verbose enable debug outputby default, the resulting segments will end up in abook/segments
The transcription tool supports up to two speakers which you can specify on the command line. The resulting voxforge-packages will end up in abook/out by default.
./abook-transcribe.py -s speaker1 -S speaker2 abook/segments/Some notes on how to segment and transcribe semi-automatically audiobooks or other audio sources (e.g. from librivox) using kaldi:
Our scripts rely on a fixed directory layout. As segmentation of librivox recordings is one of the main applications of these scripts, their terminology of books and sections is used here. For each section of a book two source files are needed: a wave file containing the audio and a text file containing the transcript.
A fixed naming scheme is used for those which is illustrated by this example:
abook/in/librivox/11442-toten-Seelen/evak-11442-toten-Seelen-1.txt abook/in/librivox/11442-toten-Seelen/evak-11442-toten-Seelen-1.wav abook/in/librivox/11442-toten-Seelen/evak-11442-toten-Seelen-2.txt abook/in/librivox/11442-toten-Seelen/evak-11442-toten-Seelen-2.wav ...
The abook-librivox.py script is provided to help with retrieval of librivox recordings and setting up the
directory structure. Please note that for now, the tool will not retrieve transcripts automatically but
will create empty .txt files (according to the naming scheme) which you will have to fill in manually.
The tool will convert the retrieved audio to 16kHz mono wav format as required by the segmentation scripts, however. If you intend to segment material from other sources, make sure to convert it to that format. For suggestions on what tools to use for this step, please refer to the manual segmentation instructions in the previous section.
NOTE: As the kaldi process is parallelized for mass-segmentation, at least 4 audio and prompt files are needed for the process to work.
This tool will tokenize the transcript and detect OOV tokens. Those can then be either replaced or added to the dictionary:
./abook-preprocess-transcript.py abook/in/librivox/11442-toten-Seelen/evak-11442-toten-Seelen-1.txtFor the automatic segmentation to work, we need a GMM model that is adapted to the current dictionary (which likely had to be expanded during transcript preprocessing) plus uses a language model that covers the prompts.
First, we create a language model tuned for our purpose:
./abook-sentences.py abook/in/librivox/11442-toten-Seelen/*.prompt
./speech_build_lm.py abook_lang_model abook abook abook parole_deNow we can create an adapted model using this language model and our current dict:
./speech_kaldi_adapt.py data/models/kaldi-generic-de-tri2b_chain-latest dict-de.ipa data/dst/lm/abook_lang_model/lm.arpa abook-de
pushd data/dst/asr-models/kaldi/abook-de
./run-adaptation.sh
popd
./speech_dist.sh -c abook-de kaldi adapt
tar xfvJ data/dist/asr-models/kaldi-abook-de-adapt-current.tar.xz -C data/models/Next, we need to create the kaldi directory structure and files for auto-segmentation:
./abook-kaldi-segment.py data/models/kaldi-abook-de-adapt-current abook/in/librivox/11442-toten-Seelennow we can run the segmentation:
pushd data/dst/speech/asr-models/kaldi/segmentation
./run-segmentation.sh
popdFinally, we can retrieve the segmentation result in voxforge format:
./abook-kaldi-retrieve.py abook/in/librivox/11442-toten-Seelen/Zamia-TTS is an experimental project that tries to train TTS voices based on (reviewed) Zamia-Speech data. Downloads here:
https://goofy.zamia.org/zamia-speech/tts/
This section describes how to train voices for NVIDIA's Tacotron 2 implementation. The resulting voices will have a sample rate of 16kHz as that is the default sample rate used for Zamia Speech ASR model training. This means that you will have to use a 16kHz waveglow model which you can find, along with pretrained voices and sample wavs here:
https://goofy.zamia.org/zamia-speech/tts/tacotron2/
now with that out of the way, Tacotron 2 model training is pretty straightforward. First step is to export filelists for the voice you'd like to train, e.g.:
./speech_tacotron2_export.py -l en -o ../torch/tacotron2/filelists m_ailabs_en mailabselliotmillernext, change into your Tacotron 2 training directory
cd ../torch/tacotron2and specify file lists, sampling rate and batch size in ''hparams.py'':
diff --git a/hparams.py b/hparams.py
index 8886f18..75e89c9 100644
--- a/hparams.py
+++ b/hparams.py
@@ -25,15 +25,19 @@ def create_hparams(hparams_string=None, verbose=False):
# Data Parameters #
################################
load_mel_from_disk=False,
- training_files='filelists/ljs_audio_text_train_filelist.txt',
- validation_files='filelists/ljs_audio_text_val_filelist.txt',
- text_cleaners=['english_cleaners'],
+ training_files='filelists/mailabselliotmiller_train_filelist.txt',
+ validation_files='filelists/mailabselliotmiller_val_filelist.txt',
+ text_cleaners=['basic_cleaners'],
################################
# Audio Parameters #
################################
max_wav_value=32768.0,
- sampling_rate=22050,
+ #sampling_rate=22050,
+ sampling_rate=16000,
filter_length=1024,
hop_length=256,
win_length=1024,
@@ -81,7 +85,8 @@ def create_hparams(hparams_string=None, verbose=False):
learning_rate=1e-3,
weight_decay=1e-6,
grad_clip_thresh=1.0,
- batch_size=64,
+ # batch_size=64,
+ batch_size=16,
mask_padding=True # set model's padded outputs to padded values
)
and start the training:
python train.py --output_directory=elliot --log_directory=elliot/logs- (1/2) Prepare a training data set
./ztts_prepare.py -l en m_ailabs_en mailabselliotmiller elliot- (2/2) Run the training
./ztts_train.py -v elliot 2>&1 | tee train_elliot.logTo build tarballs from models, use the speech-dist.sh script, e.g.:
./speech_dist.sh generic-en kaldi tdnn_sp
My own scripts as well as the data I create (i.e. lexicon and transcripts) is LGPLv3 licensed unless otherwise noted in the script's copyright headers.
Some scripts and files are based on works of others, in those cases it is my intention to keep the original license intact. Please make sure to check the copyright headers inside for more information.
- Guenter Bartsch [email protected]
- Marc Puels [email protected]
- Paul Guyot [email protected]