Aligning audio input with it’s corresponding text is a well studied research problem in speech processing.This page contains an overview of the current state of the art algorithms used for audio alignment.
An aligner is supposed to identify the time when each word in the transcription was spoken in the utterance. So ideally we would expect the following features from it:
Batch Alignment: When audio is provided in long audio files along with it’s transcription, it should be able to align them.
Live Alignment: Alignment is done on live audio, when transcription is available from before.
Precise : Aligner should be able to correct the transcription where incorrect, and then aligns it with the audio.
Error recovery : Incorrect alignment of a certain segment of audio should not effect alignment of later segments.
Memory efficient: Aligner should be careful in the amount of memory required to store and score all hypothesis under consideration, without compromising with error rate of aligner.
Disfluency detection and correction: Live audio contains skipped words, repeated words, phrases and self-corrections. Aligner should be able to detect such disfluencies and correctly align it with the correct parts of the utterance.
Minimal ASR requirement: Aligning audio data with text for a language for which we don’t have a well trained acoustic and language models.
|S.No||Algorithm||Link to paper||Author||Live Alignment||Precise||Error Recovery||MemoryEfficient||Disfluency correction||ASR required|
|2||“FST Aligner with disfluency model”||http://www.cs.columbia.edu/~julia/papers/liu03.pdf||Liu||No||Yes||No||No||Yes||Yes|
|4||“Viterbi Alignment”||http://www.cs.cmu.edu/~skishore/ksp_phdthesis.pdf||Kishore P.||Yes||No||No||Yes||No||No|
We model the speech input as a Hidden Markov Model i.e. the states of speech follow a Markov model which is unknown (hidden), and the sequence of observed features form tokens that can be used to decide which state system is in. For force aligning the audio data with a transcription, we assume that the transcription resembles the actual content in the utterance. However, this assumption is not always true. In real life data, utterances may contain disfluency or the transcription itself might not be accurate.
Disfluencies are ways in which spontaneous speech differs from written text, which includes repetitions, revisions and/or restarts in a sentence. Disfluency in an utterance is often accompanied with a change in prosodic features. Prosodic features reflect various features of the speaker or the utterance like the emotional state of the speaker; the form of the utterance (statement, question, or command); the presence of irony or sarcasm; emphasis, contrast, and focus. Liu  uses change in prosodic features as a clue for a possible disfluency in the utterance (we will call such points as interruption points).A Hidden Event Language Model is then used to look for repeated words around interruption points. If any, appropriate corrections in the alignment is made by correcting the disfluency.
If on the other hand, the transcription is inaccurate, aligner has to be designed such that it can correct such erroneous points. Hazen  and Moreno  use a speech recogniser with a n-gram language model that is strongly biased by the available transcription. Points where recogniser’s output match the transcription for a sequence of words are considered to be accurate and are marked as anchors. Identifying anchors serves the purpose of fragmenting the audio hence enabling the speech recogniser to process these fragments separately (this also has several computational advantages which we discuss later).This prevents the error of alignment of one fragment to propagate to another fragment and also reduces the number of recursions for recognition, which solves the problem of error recovery. Moreno recursively performs the same operation on fragments obtained between the anchors, and this recursion terminates only when either the transcription is completely aligned or when the duration of an unaligned audio segment is less than a pre-determined threshold. Hence this method does not correct small errors in the transcription. In each of these recursions, the language model is built specific to the words in the transcription between the anchors. Rather than recursively choosing anchors, Hazen prefers to choose anchors only once with the anchor size as small as two words. The choice of anchor size not only determines the number of anchors, but also the probability of error in it’s selection (smaller anchor size means more error).Next, a Finite state Transducer (FST) with out-of-vocabulary (OOV) filler model is used to align words between the anchors, which allows for insertion, substitution and deletion of existing words in transcription.A finite state transducer defines the set of allowed word transitions for the recogniser. Non-anchor words at this stage are marked for insertion/substitution/deletion. Lastly, the speech recogniser is re-run with full vocabulary over areas with marked words for possible insertion,substitution or deletion. Hazen’s approach hence corrects the transcription and is precise. Both Hazen and Moreno’s approach rely on availability of the full audio before alignment for anchor selection. However, live audio alignment is also an important requirement and viterbi algorithm on large audio files forms a very large beam for backtracking, which requires a lot of runtime memory making it impractical to use on a hardware with limited memory. It is observed that as the size of a viterbi lattice increases, the candidates for the optimal path mostly share a common predecessor. For example “Open Source” , “Open Suse” and “Open Sauce” might all be the hypothesis under consideration, but the word “Open” is a common predecessor in each one of these. Partial backtrack uses this observation to reduce the lattice size significantly and also gives almost live results without compromising with the recogniser’s error rate. The key to this algorithm is finding an immortal node. An immortal node is the node that terminates the common predecessor in all candidates for the optimal path.In our previous example, “Open” was an immortal node. By definition, word sequence preceding an immortal node is a part of the optimal path and hence can be recognised and taken out of the lattice. In Automatic speech and speaker recognition: Advanced topics, Kuldip Paliwal chalks out a simplified algorithm for Partial traceback. The procedure involves tracing back from the current active nodes and incrementing counter for the number of first descendants for each node encountered, until the last immortal node is reached. All dead nodes (ones with no first descendants) are deactivated after this step and the last immortal node is checked for it’s number of descendants. While this number is exactly one, it’s descendant can be treated as the new immortal node. This way we constantly reduce the size of the lattice, keeping bounds on the memory requirements for alignment. Another important and very commonly occurring issue that remains to be answered is the case when well trained acoustic model are not available for the language used in the utterance. Kishore  uses the available transcription for a first rough estimate of acoustic model and the uses Baum-Welch algorithm to re-estimate the model using another well trained acoustic model.
Recovery From Mis-Alignment
Long audio aligner in Sphinx 4 uses Out-of-Vocabulary model and grammar modifications to model dis-fluencies and to recover from mis-alignments. In real world scenarios, mis-alignments are inevitable, hence a modest objective would be to have a good recovery from such misalignment. Following is a table of experiments that can give a better understanding of optimal aligner’s variables for long audio aligner: