Alignment mass and coverage report

A corpus alignment accepts or rejects what the wider-beam retry recovers, and the overall success rate says how many utterances survived. It does not say whether what survived still covers the corpus: which speakers kept their audio, which phones and triphones rely on retries, and which ones the output lost. The mass-and-coverage report answers those questions.

The report never changes acceptance. It is built after the corpus pass, from outcomes the pass already decided. It does not change which retry-recovered alignments are accepted or rejected, the acceptance threshold, its calibration, or any default.

Where it appears

  • pstrain align prints it last, after the per-rung retry line and after any TextGrid and CTM output is written, under Mass and coverage by outcome (reporting only; acceptance is unchanged). If the report cannot be formatted, the command prints a warning in its place, and every other output is still written.

  • align_corpus attaches it to the returned job as AlignmentJob.coverage, an AlignmentCoverage. coverage.format() returns the text report and coverage.to_dict() a JSON-serializable form. Pass coverage_report=False to skip it. If the report cannot be built, for example because the dictionary cannot be read, the run logs a warning naming the error and coverage is None. The alignment itself is unaffected. When the aligner does not start, every utterance fails for that one reason, and no report is built.

  • alignment_coverage(job, transcripts, dict_path, ...) builds the report for a finished job. Use it to pass your own speaker mapping or flag parameters.

Training does not produce this report. Baum-Welch training retries failed utterances but has no acceptance check, and its failed-alignment policy is described in Failed-alignment policy.

Outcomes

Every utterance ends in exactly one outcome:

Outcome

Meaning

first pass

Aligned at the nominal beam.

retry accepted

The retry recovered it and the acceptance check accepted it. With the check off, every recovery counts here.

retry rejected

The retry recovered it and the acceptance check rejected it.

not recovered

It did not align for any other reason: the retry failed or was not run, the audio was too short for the transcript, the audio was missing, a word had no pronunciation, or the aligner did not start.

What is counted

Mass. Utterances and audio duration per outcome, overall and per speaker. Duration comes from the alignment’s frames when there is an alignment and from the WAV header otherwise. An utterance whose audio is missing or unreadable is counted with no duration, and the report says how many there were.

Speakers. pstrain has no speaker field on the alignment path. The report takes the speaker to be the text before the first / in the utterance ID (bdl/arctic_a0001 is speaker bdl), the same convention the Arctic benchmark uses. IDs without a / share one speaker, (no speaker prefix), and the text report then says no speakers are named. An ID that starts with / has an empty prefix and counts as having no speaker prefix too. Pass speaker_of= to alignment_coverage to use another mapping. The text report lists each speaker when there are 20 or fewer, and otherwise gives counts.

Phones and triphones. Token counts per outcome.

  • For first-pass and retry-accepted utterances, the phones are the alignment’s own phone segments. The base phone of a context-dependent segment is used. When phones were not captured (--no-phones), the aligned words are expanded through the dictionary, using the pronunciation variant the aligner chose.

  • For retry-rejected and not-recovered utterances there is no alignment to read. Their transcripts are expanded through the dictionary instead. Each word takes the pronunciation variant the aligned output chose most often for that word, or the dictionary’s first pronunciation when no aligned utterance contains it. A word the aligner realizes as its second variant is therefore not counted as a different unit where it failed. Words with no pronunciation are listed.

  • Filler phones (SIL and the filler dictionary’s phones) are never counted as units.

  • A triphone is written L-P+R: the phone with its neighbors among the utterance’s speech phones, across word boundaries, with SIL at the utterance edges. Fillers are removed before neighbors are taken. That includes the pauses the aligner inserts between words. Transcript expansions have no pauses, so this keeps the two on the same basis. For example, T in “at the” is AE-T+DH whether or not the speaker paused after “at”, and a failed utterance with the same words counts the same unit.

Flags

Flag

Rule

Speakers with no aligned audio

Every utterance of the speaker was rejected or not recovered.

Retry-only phones and triphones

The unit has aligned tokens only from accepted retries, and none from the first pass.

Phones and triphones seen only in failed utterances

The unit has no aligned tokens and appears only in rejected or unrecovered utterances.

Thin phones

A possible lexicon or model problem. See the rule below.

Thin-phone rule. A phone is flagged when all of these hold:

  • it has fewer than thin_tokens first-pass tokens (default 50);

  • it appears in at least thin_min_carriers utterances, in any outcome (default 3); and

  • the share of those utterances that did not align (rejected or not recovered) is at least thin_rate_ratio (default 2.0) times the run’s share of utterances that did not align. The comparison is exact and inclusive. The ratio is taken as written in decimal, so with a ratio of 3.0 and a run failure share of 10%, a phone failing 3 of its 10 utterances is flagged.

A run with no failed utterances flags no thin phones. Accepted retries do not count as failures here. A phone that aligns only through retries is flagged as retry-only instead. Flagged thin phones are listed worst first, by the share of their utterances that failed. The parameters are arguments of alignment_coverage. They are not config fields, and pstrain align uses the defaults.

Size

The text report stays bounded. Each flag list names up to 60 units and counts the rest. to_dict() names every flagged unit and has the full per-phone, per-triphone and per-speaker tables. Per-utterance outcomes are in AlignmentCoverage.outcomes.