Lattices in Kaldi

A lattice is a representation of the alternative word-sequences that are "sufficiently likely" for a particular utterance. In order to understand lattices properly you have to understand decoding graphs in the WFST framework (see Decoding graph construction in Kaldi). In this section we summarize the issues relating to Kaldi lattices, and in the rest of this page we will explain them more precisely. The reader may also want to read the paper "Generating exact lattices in the WFST framework", D. Povey, M. Hannemann et. al, ICASSP 2012 (submitted), which provides more context as to how our lattices relate to previously published algorithms. This page gives more focus to the programming aspects.

There are two representations of lattices in Kaldi. The first representation is the Lattice type. This is an FST whose weight type contains two floating point weights (the graph cost and the acoustic cost), whose input symbols are transition-ids (these are roughly like context-dependent HMM states), and whose output symbols typically represent words (in general, they represent whatever the output symbols on your decoding graph represented).

The second representation of lattices is the CompactLattice type. This contains essentially the same information as Lattice, but in a different form: it is an acceptor (meaning the input and output symbols are always identical), the input and output symbols represent words, and the sequence of transition-ids is included as part of the weights (note that OpenFst has a very general concept of weights; anything that satisfies the semiring axiom can be a weight). The weight type in CompactLattice contains a pair of floats and a sequence of integers (which represent transition-ids). We'll describe the semiring later.

Since the Lattice and CompactLattice types both represent the same information, we have made them equivalent for I/O purposes in the sense that the mechanisms that read Lattice will also read CompactLattice, and vice versa. For instance, if you use the type SequentialLatticeReader to read sets of lattices indexed by utterance-id, it will work the same whether it was actually a Lattice or a CompactLattice in the archive. However, we typically write lattices out as CompactLattice. Many programs read as Lattice because To convert between lattice types, there is a function ConvertLattice() that you can use.

The weights in both the CompactLattice and Lattice types contain two floats, which are interpreted as the graph cost and the acoustic cost respectively. The graph cost is a sum of the LM cost, the (weighted) transition probabilities, and any pronunciation cost. Since these are all mixed together, in order to do things like LM rescoring we typically first subtract the "old" LM cost and then add in the "new" LM cost, to avoid touching the other parts of the cost. If we call these two floating point values (a,b), the semiring in Lattice is like the lexicographic semiring on (a+b, a-b) (see OpenFst documentation for explanation). That is, when we "add" (in the semiring) two of these pairs, we get the pair that has the lowest (graph + acoustic) cost; if they are the same we break ties using the difference (graph - acoustic). This makes sense in terms of "taking the best path". Actually, it only makes sense if the acoustic weights are appropriately scaled. Our convention is to keep lattices externally (e.g. on disk) with the acoustic costs unscaled, and to apply the scale before any algorithm that compares the weights (and un-scale them again before we write the lattices out).

The Lattice and CompactLattice types are used as data-structures to represent conventional lattices, and also to represent N-best lists. The lattice generation algorithm we use has quite precise properties. Let lattice-delta be a number (typically 10 or so). Our lattice generation algorithm ensures (modulo beam pruning) that every word-sequence whose best alignment has cost within lattice-delta of the best path is present in the lattice, and has the best possible alignment and weight. Also, each word-sequence must be present in the lattice only once. We achieve this by first creating a state-level lattice and appropriately pruning it, and then determinizing it using a special semiring (or if you like, you can view this as a special determinization algorithm that handles non-functional FSTs by always picking the most likely alternative path). We'll explain below below exactly how this works.

We always ensure that for lattices that we write out to disk, each word-sequence has only one path through the lattice. This is required for certain algorithms (such as LM rescoring) to work properly. Being determinized (on the word labels, in the CompactLattice format) is a sufficient condition for this.

We never attach symbol tables to FSTs in any Kaldi code. In order to view archives containing lattices in their most natural form with actual words in them, we use scripts to do the conversion (an example is in egs/rm/s1/steps/decode_tri1_latgen.sh).

The Lattice type is just an FST templated on a particular semiring. In kaldi-lattice.h, we create a typedef:

typedef fst::VectorFst<LatticeArc> Lattice;

where LatticeArc is the normal arc type templated on LatticeWeight:

typedef fst::ArcTpl<LatticeWeight> LatticeArc;

LatticeWeight is again a typedef that instantiates the LatticeWeightTpl template using BaseFloat as the floating point type.

typedef fst::LatticeWeightTpl<BaseFloat> LatticeWeight;

The template LatticeWeightTpl is something that we define in the fst namespace, in fstext/lattice-weight.h. It has some similarities to the lexicographic semiring, i.e. it is similar to the OpenFst type

LexicographicWeight<TropicalWeight, TropicalWeight>

In both cases, addition (the semiring "plus" operation) is defined as taking the max, but the "max" is differently defined. LexicographicWeight first compares the first element and uses the second to break ties. LatticeWeight first compares the sum; tie-breaking is then done using the difference. So a LatticeWeight with (a,b) is equivalent to a LexigraphicWeight with (a+b, a-b). The basic intuition behind LatticeWeight is to preserve the semantics of taking the lowest-cost path (where the total cost is the graph plus acoustic cost), while separately "remembering" what the acoustic and graph cost were. In an ideal world, we might like to keep more information separate: for instance, the LM cost versus the transition-model cost versus the pronunciation-probability cost. But this is not practical because this information is all mixed together in the decoding graph, and separating it out in the decoding graph would lead to a significant expansion in decoding-graph size.

As mentioned previously, the input symbols in the Lattice represent transition-ids, and the output symbols represent words (or whatever symbol was on the output of your decoding graphs).

When designing Kaldi we considered using the LexicographicWeight type instead of the LatticeWeight type, where the first element would be the (graph + acoustic) cost and the second element would be, say, just the acoustic cost. We ultimately decided against this because while it might be slightly more efficient for some purposes, we felt it would be too confusing.

Many algorithms on lattices (for instance, taking the best path, or pruning) are most efficient to do using the Lattice type rather than the CompactLattice type. The issue is that with the CompactLattice type, the weights contain sequences of transition-ids, and operations like taking the best path will multiply weights together which corresponds to appending these sequences. For many algorithms, this would take time quadratic in the length of the lattice in words. As mentioned above, you can read a Lattice from an archive even if it contains a CompactLattice because the Holder type (LatticeHolder) does automatic conversion. So the code:

SequentialLatticeReader lattice_reader(lats_rspecifier); for (; !lattice_reader.Done(); lattice_reader.Next()) { std::string key = lattice_reader.Key(); Lattice lat = lattice_reader.Value(); ... }

would be valid even if the archive or script file specified in "lats_rspecifier" contained the CompactLattice format– which it normally would, because that is how we normally write lattices. An example of lattice type conversion is:

Lattice lat; // initialize lat. CompactLattice compact_lat; ConvertLattice(lat, &compact_lat);

Conversion to the CompactLattice type involves a "factoring" operation, the function fst::Factor() (defined by us in fstext/factor.h), that identifies chains of states that can be combined into a single CompactLattice arc. A typical example of appying an OpenFst algorithm to a lattice is as follows (this code, modified from latbin/lattice-best-path.cc, finds the best path through a lattice):

Lattice lat; // initialize lat. Lattice best_path; fst::ShortestPath(lat, &best_path);

Similarly to the Lattice type, CompactLattice is a typedef to a typical OpenFst template:

typedef fst::VectorFst<CompactLatticeArc> CompactLattice;

The CompactLatticeArc type is a standard template for an Arc, templated on CompactLatticeWeight:

typedef fst::ArcTpl<CompactLatticeWeight> CompactLatticeArc;

The CompactLatticeWeight type is a typedef as follows:

typedef fst::CompactLatticeWeightTpl<LatticeWeight, int32> CompactLatticeWeight;

The template arguments are the underlying weight type (LatticeWeight), and an integer type (int32) that is used to store the sequences of transition-ids. Note that in Kaldi code (such as above), we tend to prefer to define concrete types, for understandability, whereas the OpenFst style is to define templates, and when coding inside the fst namespace we tend to follow this convention. Thus, we define a template in the fst namespace and turn it into a fully-specified type (CompactLatticeWeight) in the kaldi namespace.

The template arguments of CompactLatticeWeightTpl are an underlying weight (here, LatticeWeight) and an integer type (here, int32). It contains two data members: a weight and a sequence of integers:

template<class WeightType, class IntType> class CompactLatticeWeightTpl { ... private: WeightType weight_; vector<IntType> string_; };

These can be accessed using the member functions Weight() and String(). The semiring used by CompactLatticeWeightTpl does not correspond to any semiring used in OpenFst, as far as we are aware. Multiplication corresponds to multiplying the weights and appending the strings together. When adding two CompactLatticeWeights, we first compare the weight component. If one of the weights is "more" than the other one, we take that weight and its corresponding string. If not (i.e. if the two weights are the same), we use an ordering on the strings to break ties. The ordering if we use is to take the shortest string if they have different length, and otherwise use lexical order (we can't just use lexical order, or it would violate one of the semiring axioms which is distributivity of multiplication over addition).

Note that we talked above about one of the weights being "more" than the other one, and we assumed that if neither is more than the other, then they are equal. This is only true if there is a total order on the semiring, and this is a requirement on the weight component of the template (this requirement is fulfilled for LatticeWeight and TropicalWeight). The "OpenFst" way to access this ordering is using the template "NaturalLess", which given a and b, looks for whether (a+b) is equal to a or to b. For efficiency reasons, we don't do it this way: we assume that there is a Compare() function defined on the weight type which returns -1, 0 or 1 depending on the comparison. We have such a function defined for the LatticeWeightTpl template. To enable CompactLatticeWeightTpl to be instantiated for TropicalWeight, which we do for testing purposes, we define an appropriate Compare function for that weight type.

There is a reason why we define this perhaps rather strange looking semiring. It relates to our idea of a lattice as containing the best path for any given word sequence. Suppose we had an arbitrary lattice (e.g., it might be a raw state-level lattice) and we represent it as a CompactLattice, which is an acceptor on word sequences with the transition-id sequences as part of the weights. The semantics of WFSTs is that for any (input-symbol-sequence, output-symbol-sequence) pair, the weight that the WFST assigns is the sum of the weights of all paths that have that pair of symbol-sequences. In the current case, we are dealing with an acceptor so there is really only symbol-sequence (the word sequence). The point is that when we sum the (weight, string) pairs in this semiring, we just get the single (weight, string) pair that has the best "weight component", i.e. the lowest cost. If we were to remove epsilons and determinize, in this semiring, we would get a lattice that only had a single path for each word sequence, and that path would have the best possible cost for that word sequence. This wouldn't be very efficient, so we actually use a specialized determinization algorithm that removes epsilons and that is optimized for this case, but it has the same effect as calling that the normal OpenFst RemoveEps() and Determinize() algorithms would.

Command-line decoding programs that have 'latgen' in their names generate lattices. Currently most of these use LatticeFasterDecoder. For purposes of exposition we will focus instead on LatticeSimpleDecoder, whose operation is simpler. This is defined in decoder/lattice-simple-decoder.h, and invoked by gmm-latgen-simple.cc. As the name suggests, LatticeSimpleDecoder is a lattice-generating decoder that is modified from SimpleDecoder. SimpleDecoder is a straightforwardly implemented Viterbi beam search algorithm with only a single tunable parameter: the pruning beam (see SimpleDecoder: the simplest possible decoder). LatticeSimpleDecoder has one more important tunable parameter: the lattice beam (or lattice-delta), which should typically be less than the pruning beam. The basic picture is that LatticeSimpleDecoder first produces a state-level lattice, and prunes it with the lattice-delta, and then at the end it runs a special determinization algorithm that results in only keeping the best path for each word sequence.

In SimpleDecoder, there are reference-counted tracebacks. In LatticeSimpleDecoder, a single traceback isn't enough because the lattice has a more complicated structure. In fact, it turned out to be more convenient to store * forward * instead of backward links. We need to do more than just do reference counting in order to free un-needed links, because of the lattice-delta pruning; in fact there is no reference counting.

The final objective of the pruning algorithm is as follows (call this the "naive" approach. Firstly, suppose that for each frame, for each state that was active on that frame we created a structure or token of some kind, and for each arc out of it (to a state that was not pruned), we create some kind of link record. Let's use the word "token" for the structure that we create for a particular graph state at a particular frame, and "link" for the forward link from one token to another token (note: these links can be within a frame as well as across frames, due to epsilon arcs in the decoding graph). The simplest possible version of the pruning algorithm is: firstly, we decode up till the end of the file, keeping all the tokens and links that were active during decoding. When we have come to the end of the file we want to apply the lattice-delta beam, such that any state or arc that is not on a path that is within lattice-delta cost of the best path, will be pruned away. This is fairly easy to do: for example, we could turn the state-level traceback into an FST and use OpenFst's Prune() algorithm. Unfortunately this wouldn't be a very practical algorithm because we would soon run out of memory: on each frame there are typically many thousands of states active, and we can't afford to wait till the end of the utterance to prune them.

Fortunately there is an efficient way to prune away these tokens without waiting until the end of the utterance. Periodically (every 25 frames by default), we do a backward sweep from the current frame to the beginning. In this backward sweep, we work out a "delta" quantity for each token. The "delta" is the difference between the cost of the best path and the cost of the best path the token is on, and it can be computed in the backward sweep (the "forward" information that we need it already inside the token, and is static). We can define an analogous "delta" quantity for the forward links. Any time this "delta" exceeds the "lattice-delta" beam, we prune away the associated token or link. For the tokens at the currently active frame, we set the "delta" to zero; we can prove that this is the right thing to do. That is, with this algorithm we will only be pruning away things that we eventually would have pruned away anyway using the "naive" algorithm.

It might seem on the face of it that our pruning algorithm would take time quadratic in the length of the utterance (because each 25 frames we visit the whole utterance up to the current point), but in fact it is more like linear, because we can detect when all the "delta" values of tokens for a particular frame have not changed, and we can stop going backward in time. We expect that for very long utterances, we will normally prune backward for only a second or two. Even if it were not for this feature, the algorithm would be dominated by a linear-time component in practical cases, because the vast majority of tokens are deleted the first time they are visited by the algorithm.

We note that there are a couple of complexities in our implementation of this. One is that because we have forward links, while the pruning algorithm goes backward in time, there is a danger of "dangling" forward links being created when we delete() tokens, so we have to avoid delete()ing tokens until we have pruned the forward links from the previous frame. The other issue is that we have to treat the final frame specially, because we need to take into account the final probabilities.

When we have reached the end of the utterance, we produce the lattice in two phases. The first phase is to create the "raw lattice". This is a state-level FST that is pruned using the lattice-delta but is not determinized, so it has many alternative paths corresponding to each word sequence. Creation of the "raw lattice" is very trivial, and consists of taking our token and link structures and translating them into OpenFst structures. If we are asked for the best path (e.g. for getting the transcript), we would just run OpenFst's ShortestPath algorithm on this raw lattice. If we are asked for the lattice, there are two options.

The first option is: the decoder supports outputting the raw state-level lattice directly (set –determinize-lattice=false); this is mostly useful if you are going to do acoustic rescoring with a system built with the same tree. Afterwards you would probably want to determinize the lattice using the lattice-determinize program, before storing the lattices on disk for an extended time (as the raw state-level lattices are quite large).

Let's assume instead that we're outputting the determinized lattice (the second option). The determinized lattice is supposed to have only one path through it for every word sequence that is allowed in the original raw lattice. We can achieve this by creating a CompactLattice, removing epsilons and determinizing, but this would be very inefficient. Instead, we run a special determinization algorithm called DeterminizeLattice(). This algorithm does epsilon removal and determinization together. In this respect it is similar to our algorithm DeterminizeStar(), but the DeterminizeLattice() algorithm uses data structures that are specifically optimized for this case. The issue is the cost overhead of appending long strings of input symbols together (the sequences of symbols are only about as long as the maximum number of frames in a word, but this is quite long). That is, if represented in the most naive way as vectors, adding a symbol to a list of N symbols and storing the result separately take time O(N). Instead we use data structures that make it take constant time. DeterminizeLattice() takes as input the Lattice format, since this is the most natural form for the input, and outputs in the CompactLattice format. However, it should really be viewed as determinization of the CompactLattice format: that is, it preserves equivalence only when viewing it as an operation on a CompactLattice.

We normally store lattices in archives. For general information about archives and Kaldi I/O mechanisms, see Kaldi I/O mechanisms. For a concrete example, a command line that generates lattices might be as follows:

gmm-latgen-simple --beam=13.0 --acoustic-scale=0.0625 exp/tri/1.mdl \ exp/graph_tri/HCLG.fst ark:test_feats.ark "ark,t:|gzip -c > exp/lats_tri/1.lats.gz"

After this, we could look at the lattice as follows:

gunzip -c exp/decode_tri2a_bg_latgen_eval92/7.lats.gz | \ utils/int2sym.pl -f 3 data/words.txt | head 444c0410 0 1 <s> 5.27832,0, 1 2 <UNK> 8.08116,1099.84, 1 3 A 12.8737,8342.93,3_1_1_1_1_1_1_1_1_1_1_12_18_17_17_17_17_17_17_17_17_17_3\ 464_3596_3632_3_1_1_1_1_1_1_1_1_1_1_1_1_1_1_1_1_1_1_1_1_1_1_1_1_1_1_1_1_1_1_1_\ 1_1_1_1_1_1_1_1_1_1_1_1_1_1_1_1_1_1_1_1_1_1_1_12_18_17_17_18562_18561_18604_18\ 603_18603_18618_604_603_603_638 1 4 IF 10.2262,8096.64,3_1_1_1_1_1_1_1_1_1_1_12_18_17_17_17_17_17_17_13708_137\ 28_13806_12654_12670_12676_3_1_1_1_1_1_1_1_1_1_1_1_1_1_1_1_1_1_1_1_1_1_1_1_1_1\ _1_1_1_1_1_1_1_1_1_1_1_1_1_1_1_1_1_1_1_1_1_1_1_1_1_1_1_1_1_12_18_17_17_17_1856\ 2 5 NOT 12.4568,10548.7,3_1_1_1_1_1_1_1_1_1_1_12_18_17_17_17_17_17_17_17_20_26\ ...

It might seem slightly disconcerting that the label <UNK> appears on a transition that has graph and acoustic scores but with no input labels (these would appear after the final comma, if they were present). It must be appreciated that graph and acoustic scores and input-label sequences are only valid when summed (or appended) over entire paths through the FST. There is no guarantee that they will be synchronized with respect to each other or with respect to the word labels.

Lattices are normally stored in the CompactLattice format in the archive, and as mentioned before, we use the convention that the acoustic weights in the archive are unscaled, so for operations (such as pruning) that are sensitive to the acoustic weight, the corresponding command-line programs would take the –acoustic-scale option and would scale the acoustic weights prior to such operations (and un-scale them afterwards).

Here we discuss some operations that can be done on lattices, and their corresponding programs.

We can prune lattices using a specified pruning beam. This will remove arcs and states that do not appear on a path that has a cost sufficiently close to the cost of the best path through the lattice. For each lattice, the program lattice-prune first scales the acoustic weights using a specified acoustic scale, then calls OpenFst's Prune() function, then reverses the scaling of acoustic weights. An example command line is:

lattice-prune --acoustic-scale=0.1 --beam=5 ark:in.lats ark:out.lats

Note: for efficiency, the pruning operation is done in the Lattice format, but we convert to CompactLattice before writing the lattice out.

The program lattice-best-path computes the best path through a lattice, and outputs the corresponding input-symbol sequence (alignment) and output-symbol sequence (transcription). As usual, the input and outputs are in the form of archives. An example command line is:

lattice-best-path --acoustic-scale=0.1 ark:in.lats ark:out.tra ark:out.ali

The program lattice-nbest computes the N best paths through the lattice (using OpenFst's ShortestPath() function), and outputs the result as a lattice (a CompactLattice), but with a special structure. As documented for the ShortestPath() function in OpenFst, the start state will have (up to) n arcs out of it, each one to the start of a separate path. Note that these paths may share suffixes. An example command line is:

lattice-nbest --n=10 --acoustic-scale=0.1 ark:in.lats ark:out.nbest

Because the "graph part" (the first component) of LatticeWeight contains the language model score mixed together with the transition-model score and any pronunciation or silence probabilities, we can't just replace it with the new language model score or we would lose the transition probabilities and pronunciation probabilities. Instead we have to first subtract the "old" LM probabilities and then add in the new LM probabilities. The central operation in both of these phases is composition (there is some scaling of weights going on, also). The command line for doing this is: first, to remove the old LM probabilities:

lattice-lmrescore --lm-scale=-1.0 ark:in.lats G_old.fst ark:nolm.lats

and to add the new LM probabilities:

lattice-lmrescore --lm-scale=1.0 ark:nolm.lats G_new.fst ark:out.lats

Note: the above examples are slightly simplified; actually you would have to "project" the G.fst FST's on the output to remove disambiguation symbols using `fstproject` with `–project_output=true`, see steps/lmrescore.sh for examples. Also there are other ways to do this; see the documentation for lattice-compose below. What the program lattice-lmrescore does is this... actually, we will first first describe the simple version that is not exactly what happens. Suppose that we are given an LM-scale S and a LM G.fst. We first multiply all the costs in G.fst by S. Then for each lattice, we compose it on the right with G, lattice-determinize it (to keep only the best path through G for each word sequence) and write it out. This would work fine for positive S, but for negative S it would be equivalent to taking the * worst * path through G, for each word sequence. To fix this, we instead do it as follows. For each input lattice, we first scale the lattice's graph (or LM) costs by the inverse of S; we then compose on the right with G.fst; we run lattice-determinization (see above) on the resulting lattice which retains only the best path through the lattice for each word sequence; and we then scale the lattice's graph/LM scores by S. This does the right thing for negative S. This whole approach only makes sense if the input lattice has only a single path through it for each word sequence (e.g. if it was lattice determinized). We assume that all lattices passed between programs have this property (this is why it is not a good idea to write out "raw" state-level lattices, unless you know what you are doing).

Note that in order for the composition to work, this program needs to map G.fst from the tropical semiring into the LatticeWeight semiring. It does this by putting the weights of G.fst into the first part of the weight (the "graph" part), and leaving the second part of the weight zero (in the semiring, this is 1). At the C++ level, this mapping is done using the MapFst mechanism of OpenFst, in which we define a Mapper class to map StdArc to LatticeArc, and then create an object of type MapFst which is evaluated on demand and converts G.fst to the LatticeWeight weight type.

It is possible to scale the probabilities on lattice weights using a general 2x2 matrix. The command-line program that does this is lattice-scale. An example command line is:

lattice-scale --acoustic-scale=0.1 --lm-scale=0.8 ark:in.lats ark:out.lats

This is actually not a frequently used program, because we expect that normally the acoustic weights in archives should be unscaled; programs that need to apply an acoustic scale generally accept an –acoustic-scale option. Such programs generally apply the scale, do an operation (such as determinization or pruning) that is sensitive to it, and then apply the inverse of the scale. For most operations we do not want to scale the LM probabilities; however, it might sometimes be useful to do so, e.g. in discriminative training it might be useful to tune both the LM and acoustic scale. The lattice-scale program accepts the options –lm2acoustic-scale and –acoustic2lm-scale. For instance, the program would add –lm2acoustic-scale times the original LM part of the weight to the acoustic part of the new weight. We include this mostly for completeness.

The program lattice-union computes the union of two lattices (as with other programs, this program iterates over sets of lattices in archives). The main envisaged use of this is in discriminative training (especially MMI), to ensure that the correct transcription is present in the denominator lattice. An example command line is:

lattice-union ark:num_lats.ark ark:den_lats.ark ark:augmented_den_lats.ark

This program calls OpenFst's Union() function, then lattice-determinizes to ensure there is only one path for each word sequence, then write the lattice out. There are actually not many situations where lattice union makes sense (e.g. it does not make much sense to do the union over lattices computed with different features or models).

The program lattice-compose works in several modes. One composes lattices with other lattices. This is done in the transducer form, i.e. viewing a lattice as a transducer from transition-ids to words. It would typically be done after first projecting one of the sets of lattices on the output (to get a word to word transducer). Typical usage might be:

lattice-project ark:1.lats ark:- | \ lattice-compose ark:- ark:2.lats ark:3.lats

Now, 3.lats would have the sum of the scores in 1.lats and 2.lats, on its paths. You can accomplish the same thing in "one package" by using the program lattice-interp (see next section).

Another mode is to compose a set of lattices with a fixed FST. The FST is converted dynamically into a lattice for this purpose; the FST weights are interpreted as the "graph part" of the lattice weights, i.e. as the first part of the tuple. An example of the use of this is to convert a lattice to phones, and then compose with a lexicon to convert back to words, for example:

lattice-to-phone-lattice final.mdl ark:1.lats ark:- | \ lattice-compose ark:- Ldet.fst ark:words.lats

This is a simplified example; you'd do further processing to remove the original "graph scores" from the lattices and add back in the grammar and transition scores. Also, Ldet.fst is a specially processed version of the lexicon: the lexicon with disambiguation symbols (but without the loop that converts #0 to #0 to pass through the language model disambiguation symbol), determinized and with the disambiguation symbols then removed.

The mode described above can also be useful for language model rescoring; however, it also comes in a different version, illustrated below (this is useful for adding in LM scores to lattices, assuming we had removed the old LM scores):

lattice-compose --phi-label=13461 ark:1.lats G.fst ark:2.lats

Here, G.fst would be the grammar including the special "backoff symbol" #0 on the input side, and 13461 is the numeric label of the backoff symbol. This form of composition treats the backoff symbol as a failure transition (using OpenFst's PhiMatcher), which is only taken in case the requested label did not exist: thus, we treat G.fst as a backoff language model. This form of composition is more exact than turning the #0 into epsilon and compsoing normally, because it prevents us taking backoff arcs when a non-backoff arc exists (which amounts to a misinterpretation of the language model). Actually, what the program lattice-compose has to do is a little more complicated than just invoking Compose with OpenFst's PhiMatcher, because of the way final-probabilities are treated. Because Compose doesn't use the matcher when computing final probabilities, we can fail to take the backoff arc when computing final probabilities. Therefore the lattice-compose program has to modify the final probabilities by "following through" the phi arcs, before calling the composition algorithm. Note that this problem would not exist if we had the end-of-sentence symbol as a label in G, but in some scripts we prefer to remove it. Note also that "phi composition" would not work properly in general, if the FST with phi's in it had epsilons on that same side. Thus, it makes sense to remove epsilons from G first.

The program lattice-interp interpolates two sets of lattices together, with a scaling factor. An example command line is:

lattice-interp --alpha=0.4 ark:1.lats ark:2.lats ark:3.lats

which amounts to putting a weight of 0.4 on the scores of 1.lats and of 0.6 on the scores of 2.lats. It keeps the alignments, if any, present in the first archive. Note that this program will not output any lattice if the composition is empty because the two lattices had no paths in common, so the final archive may have "gaps" in it. You have to build in some mechanism to produce output for such utterances, into your decoding scripts (e.g. see egs/rm/s3/steps/decode_combine.sh). The behavior of this program can be simulated by combining the programs lattice-scale, lattice-project and lattice-compose.

The program lattice-to-phone-lattice removes the word labels from the output side of a lattice and puts in their place phone labels. These are worked out from the transition-ids on the input side. Note that the phone labels are not exactly aligned with the phone boundaries (as a rule, FSTs have no real concept of alignment between input and output symbols). Typical usage is:

lattice-to-phones final.mdl ark:1.lats ark:phones.lats

Projection is an FST operation that turns an FST into an acceptor, either by copying the output symbols to the input, or the input symbols to the output, so they are both the same. The default for lattice-project is to copy the word labels to the input side (this is useful when doing interpolation of the scores of lattices). An example is:

lattice-project ark:1.lats ark:- | \ lattice-compose ark:- ark:2.lats ark:3.lats

The program lattice-equivalent is mostly useful as a debugging tool. It tests the equivalence of sets of lattices. It returns with exit status 0 only if all pairs of lattices in two supplied tables (e.g. archives) are equivalent. It uses a randomized equivalence testing algorithm to do this. A typical usage is:

lattice-equivalent ark:1.lats ark:2.lats || echo "Not equivalent!"

Optional arguments include –num-paths (the number of paths used in the randomized testing algorithm) and –delta, the score difference at which scores are considered different for purposes of equivalence testing.

The program lattice-rmali removes the alignment information (i.e. the transition-ids) from the input side of the lattices. This is mainly useful if you will not need that information (e.g. you only need the lattices for language model rescoring), and you want to save disk space. An example is:

lattice-rmali ark:in.lats ark:word.lats

The program lattice-boost-ali is used in boosted MMI training. It reads in a lattice and an alignment (both as tables, e.g. archives), and outputs the lattices with the language model scores (i.e. the graph part of the scores) boosted by the parameter b times the number of frame-level phone errors on that path. Typical usage is:

lattice-boost-ali --silence-phones=1:2:3 --b=0.1 final.mdl ark:1.lats \ ark:1.ali ark:boosted.lats

The silence phones are treated specially: wherever they appear in the lattices, they are assigned zero error, i.e. they are not counted as errors (this behavior can be controlled with the –max-silence option). Note that this special treatment of silence has been adopted from MPE training where it appeared to help; we have not investigated its exact effect on boosted MMI.

The program lattice-to-post computes, from a lattice, per-frame posterior probabilities of the transition-ids in the lattice. This is done by a standard forward-backward type of algorithm. Since, by convention, we store lattices without any acoustic scaling, it will normally be necessary to supply an acoustic scale to be used in the forward-backward algorithm. It also accepts a language model scale, but the default (1.0) will often be most appropriate. An example of using this program is:

lattice-to-post --acoustic-scale=0.1 ark:1.lats ark:- | \ gmm-acc-stats 10.mdl "$feats" ark:- 1.acc

The program lattice-determinize implements lattice-determinization, which essentially consists of keeping, for each word-sequence in the lattice, only the best-scoring sequence of transition-ids. In general this process is sensitive to the acoustic scale (because "best-scoring" depends on the scale), but if the lattice has previously been determinized and then has only been processed in a "word-level" way, i.e. each word-sequence still has only one transition-id sequence, then the acoustic scale doesn't matter. The only time the acoustic scale is likely to matter is if you generate state-level lattices, e.g. you do lattice generation with –determinize-lattice=false. This program has other options, that are mostly related to what to do if determinization "blows up", i.e. exhausts memory, but in general you can just leave these at their default values. These are mostly there because a previous version of the determinization tended to exhaust memory. You may want to set –prune=true if you want to do pruning as part of the same program, but be careful that this only makes sense if you have set the acoustic scale to an appropriate value; you might also want to set the –beam option in that case.

A typical usage is:

lattice-determinize ark:1.lats ark:det.lats

or:

lattice-determinize --acoustic-scale=0.08333 ark:state_level.lats ark:1.lats

Note that lattices produced by lattice generation programs will be default already be determinized, so it only makes sense to do determinization if you have just done an operation like composition that makes lattices nondeterministic, or if you have given –determinize-lattice=false to the lattice generation program to create state-level lattices.

The program lattice-oracle takes two tables: the first of lattices, the second of transcriptions (C++ representation: vector<int32>), and outputs the oracle word-sequence of the lattices; it prints out the corresponding WERs in its logging output. This is done by constructing an "edit-distance FST" and composing this with unweighted acceptors constructed from the lattice and from the reference.

An example (this script fragment assumes the newer, s3-style script) is:

cat $data/text | \ sed 's:<NOISE>::g' | sed 's:<SPOKEN_NOISE>::g' | \ scripts/sym2int.pl --ignore-first-field $lang/words.txt | \ lattice-oracle --word-symbol-table=$lang/words.txt \ "ark:gunzip -c $dir/lats.pruned.gz|" ark:- ark,t:$dir/oracle.tra \ 2>$dir/oracle.log

The program lattice-add-trans-probs adds in the (negated) transition probabilities to the costs of a lattice. This is useful in forms of lattice rescoring where the original "graph costs" are discarded and reconstructed from scratch. A typical usage is:

lattice-add-trans-probs --transition-scale=1.0 --self-loop-scale=0.1 \ final.mdl ark:1.lats ark:2.lats

See Scaling of transition and acoustic probabilities for more explanation about these scaling factors.

The program lattice-to-fst converts an table of lattices into a table of FSTs, in the same format used by the code that compiles the training graphs. The resulting FSTs will be weighted word acceptors (although in current use of this program we apply a scale of zero so there are actually no weights). These may be used as input to the program compile-train-graphs-fsts to produce decoding graphs that only include the paths present in the lattice. The main use of this is to do lattice rescoring across different models that don't share the same tree. An example of usage is:

lattice-to-fst --lm-scale=0.0 --acoustic-scale=0.0 ark:1.lats ark:1.words

The program lattice-copy simply copies a table (e.g. archive) of lattices. This is mostly useful if you want to view binary FSTs in text form. For example:

lattice-copy ark:1.lats ark,t:- | head -50

There are some situations (e.g. Viterbi training; rescoring with neural-net language models) where we don't want the lattice structure and want either the best path or the N best paths. Our format for N-best lists is the same as for lattices, except for each utterance we have multiple FSTs (up to n, for some specified n). Suppose the utterance id's are uttA, uttB, and so on. A normal Table of lattices (e.g. an archive) will contain a lattice for uttA, one for uttB, and so on. If we run the program

lattice-to-nbest --acoustic-scale=0.1 --n=10 ark:1.lats ark:1.nbest

then the archive 1.nbest will contain lattices for uttA-1, uttA-2, ... uttA-10, for uttB-1 ... uttB-10, and so on. Of course some of the utterances may not have that many N-best entries if the lattice did not contain that many distinct word sequences. The program lattice-to-nbest needs the acoustic scale as this affects which paths would have the lowest score.

For some purposes it will still be inconvenient to deal with an FST and we want an actual linear structure, e.g. just a sequence of words. For this, we can use the following program:

nbest-to-linear ark:1.nbest ark:1.ali ark:1.words ark:1.lmscore ark:1.acscore

This program reads in an archive of lattices, which must each be a linear FST, and outputs 4 archives, corresponding to the input symbol sequences, the output symbol sequences, and the acoustic and LM scores. The format (assuming you write in text mode, e.g. with "ark,t:my_file") is as one would expect, e.g.:

# head 1.words utt-A-1 10243 432 436 10244 utt-A-2 7890 420 10244 ...

Remember that Kaldi always deals with integer id's; you can convert to words using the symbol table and "utils/int2sym.pl -f 2-". The reverse transformation can be accomplished with linear-to-nbest.

A program that might also be useful is lattice-to-1best, which is not the same as lattice-to-nbest with –n=1, because it doesn't append "-1" to the utterance id. You can pipe the output of this into nbest-to-linear if you want the word sequences or something of that nature. The program lattice-best-path takes in a lattice and outputs up to two archives, containing the word sequences and transition-id sequences. This is there for historical reasons; the combination of lattice-to-1best and nbest-to-linear is cleaner and more consistent with the overall framework.

If you want the time information of the lattices explicitly (rather than having to count transition-ids), there is a function LatticeStateTimes (for Lattice), and CompactLatticeStateTimes (for CompactLattice), which will give you the time where each state is located (a number from 0 to the number of frames in the file). Be careful that in general, the words are not synchronized with to the transition-ids, meaning that the transition-ids on an arc won't necessarily all belong to the word whose label is on that arc. This means that the times you get from the lattice will (as far as the word labels are concerned) be inexact. The same is also true of the weights; these are also not synchronized with either the words or the transition-ids on a particular arc. If you want exact times (e.g. for conversion to HTK lattices, or for sclite scoring), then you should run the program lattice-align-words. This program only works if you built your system with word-position-dependent phones, and it requires certain command-line options to tell it which phones are in which position in the word. See egs/wsj/s3/run.sh for an example (search for align). There is an alternative program, lattice-align-words-lexicon, that you can use if your system does not have word-position-dependent phones.