GStreamer plugin that wraps Kaldi's SingleUtteranceNnet2Decoder. It requires iVector-adapted DNN acoustic models. The iVectors are adapted to the current audio stream automatically.
2017-08-09: Also implemented phone confidence scores, which are included with phone alignments. Note that phone confidence scores are very inaccurate.
2017-08-09: Implemented word confidences. Confidence scores are included with word alignments,
so set word-boundary-file=dir/word_boundary.int, since this invokes word alignment generation.
2017-06-05: Now works with Kaldi 5.2 (i.e., the latest), also requires it. Decoder now also sets test mode for dropout and batch norm layers, and uses more memory-efficient FST representation for reading the decode graph (HCLG.fst).
2017-02-23: Now works with the latest Kaldi (also requires very fresh Kaldi). This also means that LSTM nnet3 models (including chain models) are supported (not tested).
2016-10-14: Support for nnet3 (including 'chain') models, thanks to @yifan! Not tested very carefully.
Set the nnet-mode property to 3 to use nnet3 models. Note that this property must be set before the other
properties.
2016-06-23: Got rid of the GStreamer property warnings at startup, thanks to @MathieuDuponchelle.
2015-11-05: Minor changes for compatibility with Ubuntu 12.04 (and its older version
of Jansson). Also use ReadDecodeGraph to read fst file, supporting the use of
const (mapped) fsts.
2015-06-03: Full results can now inlude word alignment information for the
best hypothesis. In order to activate this, set the word-boundary-file
property to point to the word_boundary.int file typically found in the
data/lang/phones/ subdirectory of a trained system.
2015-06-02: Added functionality to output structured recognition results,
that optionally include n-best results, phone alignment information and
possibly other things (word alignments, sausages, lattices, sentence and
word confidence scores) in the future.
The structured results are pushed using the signal full-final-result, and
are formatted using a JSON encoding. See below for a full description.
Also removed the individual signals that were used for pushing out phone alignments.
2015-04-30: Added functionality to change models (FST, acoustic model,
big language model) after initial initialization. Added functionality to
ouput phone aligment information (see properties do-phone-alignment
and phone-syms). Both additions by Ricard Marxer (@rikrd).
2015-04-28: Endpointing and interim recognition results now work correctly
when using the threaded decoder. Note that endpointing does not work exactly
the same as with the unthreaded decoder, and there might be some differences
between individual decoder runs, but for all practical purposes it should be OK.
Also, introduced a new property traceback-period-in-secs
that specifies how often intermediate results are sent to the client (default is 0.5).
NB: this update requires Kaldi revision 5036 or later.
2015-03-05: Threaded decoder can now be selected at configuration time, using the
use-threaded-decoder property. NB: this property should be set before other
properties. Endpointing and partial results might still not work as expected with the threaded decoder.
2015-03-01: one can now optionally use the threaded online decoder. Citing
Daniel Povey who developed the code in Kaldi: "This should make it possible to
decode in real-time with larger models and graphs than before, because
the decoding and the nnet evaluation are in separate threads and can be done in parallel."
Use make CPPFLAGS=-DTHREADED_DECODER to compile it. Note that the endpointing
and partial results might not work as expected with the threaded decoder,
working on the fix.
2015-01-09: Added language model rescoring functionality. In order to use it,
you have to specify two properties: lm-fst and big-lm-const-arpa. The lm-fst
property gives the location of the original LM (the one that was used fpr
compiling the HCLG.fst used during decoding). The big-lm-const-arpa property
gives the location of the big LM used that is used to rescore the final lattices.
The big LM must be in the 'ConstArpaLm' format, use the Kaldi's
utils/build_const_arpa_lm.sh script to produce it from the ARPA format.
2014-11-11: the plugin saves the adaptation state between silence-segmented utterances and between
multiple decoding sessions of the same plugin instance.
That is, if you start decoding a new stream, the adaptation state of the
previous stream is used (unless it's the first stream, in which case a global mean is used).
Use the adaptation-state plugin property to get, and set the adaptation state. Use an empty string
with the set method to reset the adaptation state to default. This functionality requires
Kaldi revision 4582 or later.
The following works on Linux (I'm using Debian 'testing'; Ubuntu 12.04 also works).
Compile Kaldi trunk, using the shared configuration: In Kaldi's 'src' directory:
./configure --shared
make depend
make
Install gstreamer-1.0:
sudo apt-get install gstreamer1.0-plugins-bad gstreamer1.0-plugins-base gstreamer1.0-plugins-good gstreamer1.0-pulseaudio gstreamer1.0-plugins-ugly gstreamer1.0-tools libgstreamer1.0-dev
On Ubuntu 12.04, you'll need to use a backport ppa to get version 1.0 of gstreamer, before doing the sudo apt-get install above:
sudo add-apt-repository ppa:gstreamer-developers/ppa
sudo apt-get update
Install the Jansson library development package (version 2.7 or newer), used for encoding results as JSON:
sudo apt-get install libjansson-dev
Now we can compile this plugin. Change to src of this project:
cd src
Compile, specifying Kaldi's root directory:
make depend
KALDI_ROOT=/path/of/kaldi-trunk make
This should result in 'libgstkaldionline2.so'.
Test if GStreamer can access the plugin:
GST_PLUGIN_PATH=. gst-inspect-1.0 kaldinnet2onlinedecoder
The output should list all plugin properties with their default values:
Factory Details:
Rank none (0)
Long-name KaldiNNet2OnlineDecoder
Klass Speech/Audio
Description Convert speech to text
[...]
name : The name of the object
flags: readable, writable
String. Default: "kaldinnet2onlinedecoder0"
parent : The parent of the object
flags: readable, writable
Object of type "GstObject"
silent : Silence the decoder
flags: readable, writable
Boolean. Default: false
model : Filename of the acoustic model
flags: readable, writable
String. Default: "final.mdl"
[...]
max-nnet-batch-size : Maximum batch size we use in neural-network decodable object, in cases where we are not constrained by currently available frames (this will rarely make a difference)
flags: readable, writable
Integer. Range: -2147483648 - 2147483647 Default: 256
Element Signals:
"partial-result" : void user_function (GstElement* object,
gchararray arg0,
gpointer user_data);
"final-result" : void user_function (GstElement* object,
gchararray arg0,
gpointer user_data);
Command-line usage is demonstrated in demo/.
Usage through GStreamer's Python bindings is demonstrated demo/gui-demo.py
and in the project
https://github.com/alumae/kaldi-gstreamer-server, in file kaldigstserver/decoder2.py.
Below is a sample of JSON-encoded full recognition results, pushed out
using the full-final-result signal. This sample was generated using
do-phone-alignment=true and num-nbest=10 (although due to pruning it
includes only two n-best hypotheses). Note that the words in the file specified
by word-syms and the phones in the file specified in phone-syms must
be encoded using UTF-8, otherwise the output won't be valid JSON.
{
"status": 0,
"result": {
"hypotheses": [
{
"transcript": "we're not ready for the next epidemic",
"likelihood": 128.174,
"phone-alignment": [
{
"start": 0,
"phone": "SIL",
"length": 0.71
},
{
"start": 0.71,
"phone": "W_B",
"length": 0.18
},
{
"start": 0.89,
"phone": "ER_E",
"length": 0.06
},
{
"start": 0.95,
"phone": "N_B",
"length": 0.06
},
{
"start": 1.01,
"phone": "AA_I",
"length": 0.19
},
{
"start": 1.2,
"phone": "T_E",
"length": 0.11
},
{
"start": 1.31,
"phone": "R_B",
"length": 0.07
},
{
"start": 1.38,
"phone": "EH_I",
"length": 0.1
},
{
"start": 1.48,
"phone": "D_I",
"length": 0.05
},
{
"start": 1.53,
"phone": "IY_E",
"length": 0.22
},
{
"start": 1.75,
"phone": "SIL",
"length": 0.46
},
{
"start": 2.21,
"phone": "F_B",
"length": 0.1
},
{
"start": 2.31,
"phone": "ER_E",
"length": 0.05
},
{
"start": 2.36,
"phone": "DH_B",
"length": 0.05
},
{
"start": 2.41,
"phone": "AH_E",
"length": 0.05
},
{
"start": 2.46,
"phone": "N_B",
"length": 0.06
},
{
"start": 2.52,
"phone": "EH_I",
"length": 0.11
},
{
"start": 2.63,
"phone": "K_I",
"length": 0.08
},
{
"start": 2.71,
"phone": "S_I",
"length": 0.05
},
{
"start": 2.76,
"phone": "T_E",
"length": 0.07
},
{
"start": 2.83,
"phone": "EH_B",
"length": 0.08
},
{
"start": 2.91,
"phone": "P_I",
"length": 0.09
},
{
"start": 3,
"phone": "AH_I",
"length": 0.04
},
{
"start": 3.04,
"phone": "D_I",
"length": 0.08
},
{
"start": 3.12,
"phone": "EH_I",
"length": 0.1
},
{
"start": 3.22,
"phone": "M_I",
"length": 0.08
},
{
"start": 3.3,
"phone": "IH_I",
"length": 0.08
},
{
"start": 3.38,
"phone": "K_E",
"length": 0.18
},
{
"start": 3.56,
"phone": "SIL",
"length": 0.13
}
],
"word-alignment": [
{
"word": "we're",
"start": 0.71,
"length": 0.24
},
{
"word": "not",
"start": 0.95,
"length": 0.36
},
{
"word": "ready",
"start": 1.31,
"length": 0.44
},
{
"word": "for",
"start": 2.21,
"length": 0.15
},
{
"word": "the",
"start": 2.36,
"length": 0.1
},
{
"word": "next",
"start": 2.46,
"length": 0.37
},
{
"word": "epidemic",
"start": 2.83,
"length": 0.73
}
]
},
{
"transcript": "were not ready for the next epidemic",
"likelihood": 125.323
}
]
},
"segment-start": 58.25,
"segment-length": 3.69,
"total-length": 61.94
}
If you use this software for research, you can cite the following paper (available here: http://ebooks.iospress.nl/volumearticle/37996):
@inproceedigs{alumae2014,
author={Tanel Alum\"{a}e},
title="Full-duplex Speech-to-text System for {Estonian}",
booktitle="Baltic HLT 2014",
year=2014,
address="Kaunas, Lithuania"
}
Of course, you should also acknowledge Kaldi, which does all the hard work.