Skip to content

Fix two-pass state alignment not reaching the final state #468

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
lenzo-ka wants to merge 4 commits into
base: main
Choose a base branch
from
Open

Fix two-pass state alignment not reaching the final state #468

Changes from all commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
850a642
Don't advance active successor's frame in phone_transition
lenzo-ka Apr 23, 2026
ceee17c
Fix ef boundary handling in state_align_search
lenzo-ka Apr 23, 2026
1c18c6a
refactor: remove a stray space
dhdaines May 4, 2026
a19653f
Merge remote-tracking branch 'origin/main' into kal-align-2pass
dhdaines May 6, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
34 changes: 24 additions & 10 deletions src/state_align_search.c
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -94,10 +94,12 @@ prune_hmms(state_align_search_t *sas, int frame_idx)
hmm_t *hmm = sas->hmms + i;
if (hmm_frame(hmm) < frame_idx)
continue;
/* Enforce alignment constraint: due to non-emitting states,
* previous phone's HMM remains active in first frame of its
* successor. */
if (nf > sas->ef[i])
/* ef[i] is the exclusive end frame (= start + duration).
* Once nf reaches ef[i] the phone has used its full window;
* stop advancing so it is not scored in the next frame.
* phone_transition is allowed to fire at this last active
* frame via the complementary check in phone_transition. */
if (nf >= sas->ef[i])
continue;
hmm_frame(hmm) = nf;
}
Expand All @@ -114,18 +116,30 @@ phone_transition(state_align_search_t *sas, int frame_idx)
int32 newphone_score;

hmm = sas->hmms + i;
if (hmm_frame(hmm) != nf)
/* Allow transition from phones that were active this frame,
* whether they were advanced to nf (normal case) or left at
* frame_idx by prune_hmms because they hit their ef boundary.
* The original check (hmm_frame != nf) silently blocked
* cross-word transitions when the phone was evaluated at
* exactly its last allowed frame. */
if (hmm_frame(hmm) < frame_idx)
continue;
/* Enforce alignment constraint for initial state of each phone. */
if (nf < sas->sf[i + 1])
continue;

newphone_score = hmm_out_score(hmm);
/* Transition into next phone using the usual Viterbi rule. */
nhmm = hmm + 1;
if (hmm_frame(nhmm) < frame_idx
|| newphone_score BETTER_THAN hmm_in_score(nhmm)) {
if (hmm_frame(nhmm) < frame_idx) {
/* Successor is inactive: enter it fresh with the new frame. */
hmm_enter(nhmm, newphone_score, hmm_out_history(hmm), nf);
continue;
}
/* Successor is already active: update score/history only (Viterbi
* competition), never bump its hmm_frame to a later value. */
if (newphone_score BETTER_THAN hmm_in_score(nhmm)) {
hmm_in_score(nhmm) = newphone_score;
hmm_in_history(nhmm) = hmm_out_history(hmm);
}
}
}
Expand Down Expand Up @@ -243,7 +257,7 @@ state_align_search_finish(ps_search_t *search)
ent = ps_alignment_iter_get(itor);
ent->start = cur_frame + 1;
ent->duration = last_frame - ent->start;
ent->score = last.score - cur.score;
ent->score = last.score - cur.score;
E_DEBUG("state %d start %d end %d\n", last.id,
ent->start, last_frame);
last = cur;
Expand Down Expand Up @@ -458,7 +472,7 @@ state_align_search_init(const char *name,
++i, itor = ps_alignment_iter_next(itor)) {
ps_alignment_entry_t *ent = ps_alignment_iter_get(itor);
int min_nframes;

hmm_init(sas->hmmctx, &sas->hmms[i], FALSE,
ent->id.pid.ssid, ent->id.pid.tmatid);
/* Can't align less than the number of frames in an HMM! */
Expand Down
Loading