1 Introduction
1In the last five years ASR technology has achieved remarkable results, thanks to increased training data, computational resources, and the use of deep neural networks (DNNs, (LeCun et al., 2015)). However, the performance of connectionist ASR degrades when testing conditions are different from training conditions (e.g., acoustic environments are different) (Huang et al., 2014). This is usually referred to as the training-testing mismatch problem. This problem is partly masked by multi-condition training (Seltzer et al., 2013) that consists in using very large training datasets (up to thousands of hours) of transcribed speech to cover as many sources of variability as possible (e.g., speaker’s gender, age and accent, different acoustic environments).
2One of the two main goals of the ArtiPhon task at Evalita 2016 was to evaluate phone recognition systems in mismatched speaking styles. Between training and testing data the speaking style was the condition that differed. More specifically, while the training dataset consists of read speech where the speaker was required to keep a constant speech rate, testing data range from slow and hyper-articulated speech to fast and hypo-articulated speech. Training and testing data are from the same speaker.
3The second goal of the ArtiPhon task was to investigate whether the use of speech production knowledge, in the form of measured articulatory movements, could help in building ASR systems that are more robust to the effects of the mismatch problem.
4The use of speech production knowledge, i.e., knowledge about how the vocal tract behaves when it produces speech sounds, is motivated by the fact that complex phenomena observed in speech, for which a simple purely acoustic description has still to be found, can be easily and compactly described in speech production-based representations. For example, in Articulatory Phonology (Browman and Goldstein, 1992) or in the distinctive features framework (Jakobson et al., 1952) coarticulation effects can be compactly modeled as temporal overlaps of few vocal tract gestures. The vocal tract gestures are regarded as invariant, i.e., contextand speaker-independent, production targets that contribute to the realization of a phonetic segment. Obviously the invariance of a vocal tract gesture partly depends on the degree of abstraction of the representation but speech production representations offer compact descriptions of complex phenomena and of phonetic targets that purely acoustic representations are not able to provide yet (Maddieson, 1997).
5Recently, my colleagues and I have proposed DNN-based “articulatory” ASR where the DNN that computes phone probabilities is forced, during training, to learn/use motor features. We have proposed strategies that allow motor information to produce an inductive bias on learning. The bias resulted in improvements over strong DNN-based purely auditory baselines, in both speaker-dependent (Badino et al., 2016) and speaker-independent settings (Badino, 2016)
6Regarding the Artiphon task, unfortunately only one out of the 6 research groups that expressed an interest in the task actually participated (Piero Cosi from ISTC at CNR, henceforth I will refer to this participant as ISTC) (Cosi, 2016). The ISTC system did not use articulatory data.
7In this report I will present results of the ISTC phone recognition systems and of baseline systems that also used articulatory data.
2 Data
8The training and testing datasets used for the ArtiPhon task were selected from voice cnz of the Italian MSPKA corpus (http://www.mspkacorpus.it/) (Canevari et al., 2015), which was collected in 2015 at the Istituto Italiano di Tecnologia (IIT).
9The training dataset corresponds to the 666- utterance session 1 of MPSKA, where the speaker was required to keep a constant speech rate. The testing dataset was a 40-utterance subset selected from session 2 of MPSKA. Session 2 of MPSKA contains a continuum of ten descending articulation degrees, from hyper-articulated to hypo-articulated speech. Details on the procedure used to elicit this continuum are provided in (Canevari et al., 2015).
10Articulatory data consist of trajectories of 7 vocal tract articulators and recorded with the NDI (Northern Digital Instruments, Canada) wave speech electromagnetic articulography system at 400 Hz.
11Seven 5-Degree-of-freedom (DOF) sensor coils were attached to upper and lower lips (UL and LL), upper and lower incisors (UI and LI), tongue tip (TT), tongue blade (TB) and tongue dorsum (TD). For head movement correction a 6-DOF sensor coil was fixed on the bridge of a pair of glasses worn by the speakers.
12The NDI system tracks sensor coils in 3D space providing 7 measurements per each coil: 3 positions (i.e. x; y; z) and 4 rotations (i.e. Q0;Q1;Q2;Q3) in quaternion format with Q0 = 0 for 5-DOF sensor coils.
13Contrarily to other articulographic systems (e.g. Carstens 2D AG200, AG100) speakers head is free to move. That increases comfort and the naturalness of speech.
14During recordings speakers were asked to read aloud each sentence that is prompted on a computer screen. In order to minimize disfluencies speakers had time to silently read each sentence before reading out.
15The audio files of the MSPKA corpus are partly saturated.
16The phone set consists of 60 phonemes, although the participants could collapsed them to 48 phonemes as proposed in (Canevari et al., 2015).
3 Sub-tasks
17In Artiphon sub-tasks are phone recognition tasks. The participants were asked to:
-
train phone recognition systems on the training dataset and then run them on the test dataset;
-
(optional) use articulatory data to build “articulatory” phone recognition systems.
18Articulatory data were also provided in the test dataset thus three different scenarios were possible:
-
Scenario 1. Articulatory data not available
-
Scenario 2. Articulatory data available at both training and testing.
-
Scenario 3. Articulatory data available only at training.
19Note that only scenarios 1 and 3 are realistic ASR scenarios as during testing articulatory data are very difficult to access.
20Participants could build purely acoustic and articulatory phone recognition systems starting from the Matlab toolbox developed at IIT, available at https://github.com/robotology/natural-speech.
4 Phone recognition systems
21Baseline systems are hybrid DNN-HMM systems while ISTC systems are either GMM-HMM or DNN-HMM systems with DNN-HMM.
22The ISTC systems were trained using the KALDI ASR engine. ISTC systems used either the full phone set (with 60 phone labels) or a reduced phone set (with 29 phones). In the reduced phone set all phones that are not actual phonemes in current Italian were correctly removed. However, important phonemes were also arbitrarily removed, most importantly, geminates and corresponding non-geminate phones were collapsed into a single phone (e.g., /pp/ and /p/ were both represented by label /p/).
23ISCT systems used either monophones or triphones.
24ISCT systems were built using KALDI (Povey et al., 2011) with TIMIT recipes adapted to the APASCI dataset (Angelini & al., 1994). Two training datasets where used:
-
the single-speaker dataset provided within the ArtiPhon task;
-
the APASCI dataset.
25In all cases only acoustic data were used (scenario 1), so the recognition systems were purely acoustic recognition systems. Henceforth I will refer to ISTC systems trained on the ArtiPhon single-speaker training dataset as speaker-dependent ISTC systems (as the speaker in training and testing data is the same) and to ISTC systems trained on the APASCI dataset as speaker-independent ISTC systems. Baseline systems were built using the aforementioned Matlab toolbox and only trained on the ArtiPhon training dataset (so they are all speaker-dependent systems).
26Baseline systems used a 48 phone set and three-state monophones (Canevari et al., 2015). Baseline systems were trained and tested according to all three aforementioned three scenarios. The articulatory data considered only refer to x-y positions of 6 coils (the coil attached to the upper teeth was exluded).
5 Results
27Here I report some of the most relevant results regarding ISTC and baseline systems.
28Baseline systems and ISTC systems are not directly comparable as very different assumptions were made, most importantly they use different phone sets.
29Additionally, ISCT systems were mainly concerned with exploring the best performing systems (created using well-known KALDI recipes for ASR) and comparing them in the speaker-dependent and in the speakerindependent case.
30Baselines systems were created to investigate the utility of articulatory features in mismatched speaking styles.
5.1 ISCT systems
31Here I show results on ISCT systems trained and tested on the 29 phone set. Table 1 shows results of the speaker-dependent systems while Table 2 shows results in the speakerindependent case.
32The results shown in the two tables refer to the various training and decoding experiments, see (Rath et al., 2013) for all acronyms references:
-
MonoPhone (mono);
-
Deltas + Delta-Deltas (tri1);
-
LDA + MLLT (tri2);
-
LDA + MLLT + SAT (tri3);
-
SGMM2 (sgmm2_4);
-
MMI + SGMM2 (sgmm2_4_mmi_b0.1-4);
-
Dan’s Hybrid DNN (tri4-nnet),
-
system combination, that is Dan’s DNN + SGMM (combine_2_1-4);
-
Karel’s Hybrid DNN (dnn4_pretrain-dbn_dnn);
-
system combination that is Karel’s DNN + sMBR (dnn4_pretrain-dbn_dnn_1-6).
Table 1. Results of ISCT systems on speaker-dependent sub-task with the 29-phone set. MFCC: Mel-Frequency Cepstral Coefficients; LDA: Linear Discriminant Analysis; MLTT: Maximum Likelihood Linear Transform; fMLLR: feature space Maximum Likelihood Linear Regression; CMN: Cepstral Mean Normalization. MMI: Maximum Mutual Information; BMMI: Boosted MMI; MPE: Minimum Phone Error; sMBR: State-level Minimum Bayes Risk
Table 2. Results of ISCT systems on speaker independent sub-task with 29-phone set.
Figure 1. Phone Error Rate (PER) over 10 degrees of articulation when (i) using only MFSCs (black), (ii) using MFSCs appended to actual VTFs (light grey) and (iii) using MFSCs appended to recovered VTFs (dark grey)
33The most interesting result is that while DNN-HMM systems outperform GMM-HMM systems in the speaker-independent case (as expected), GMM-HMM and more specifically sub-space GMM-HMM (Povey et al., 2011), outperform the DNN-based systems in the speaker dependent case.
34Another interesting result is that sequence based training strategies (Vesely et al., 2013) did not produce any improvement over framebased training strategies.
5.2 Baseline systems – acoustic vs. articulatory results
35The baseline systems addressed the two main questions that motivated the design of the ArtiPhon task: (i) how does articulatory information contribute to phone recognition?; (ii) how does the phone recognition system performance vary along the continuum from hyper-articulated speech to hypo-articulated speech?
36Figure 1 shows phone recognition results of 3 different systems over the 10 degrees of articulation from hyper-articulated to hypoarticulated speech.
37The three systems, reflecting the three aforementioned training-testing scenarios, are:
-
phone recognition system that only uses acoustic feature, specifically mel-filtered spectra coefficients (MFSCs, scenario 1)
-
articulatory phone recognition system where actual measured articulatory/vocal tract features (VTFs) are appended to the input acoustic vector during testing (scenario 2)
-
articulatory phone recognition system where reconstructed VTFs are appended to the input acoustic vector during testing (scenario 3)
38The last system reconstructs the articulatory features using an acoustic-to-articulatory mapping learned during training (see, e.g., (Canevari et al., 2013)for details).
39All systems used a 48-phone set as in cc.
40One first relevant result is that all systems performed better at high levels of hyperarticulation than at “middle” levels (i.e., levels 5-6) which mostly corresponds to the training condition (Canevari et al., forthcoming). In all systems performance degraded from hyperto hypo-articulated speech.
41Reconstructed VTFs always decrease the phone error rate. Appending recovered VTFs to the acoustic feature vector produces a relative PER reduction that ranges from 4.6% in hyper-articulated speech, to 5.7% and 5.2% in middle- and hypo-articulated speech respectively.
42Actual VTFs provide a relative PER reduction up to 23.5% in hyper-articulated speech, whereas, unexpectedly, no improvements are observed when actual VTFs are used in middle- and hypo-articulated speech. That might due to the fact that sessions 1 and 2 of the MSPKA corpus took place in different days so EMA coils could be in slightly different positions.
6 Conclusions
43This paper described the ArtiPhon task at Evalita 2016 and showed and discussed results of baseline phone recognition systems and of the submitted systems.