CtcTrellis.Align accepts row-major natural-log frame probabilities and a fixed
sequence of nonblank vocabulary IDs. It consumes every frame and requires the
complete target. Leading and trailing blank are allowed. Blank separates equal
adjacent target IDs; distinct adjacent IDs may follow without blank. The recurrence
is the blank-expanded target in Graves et al., ICML 2006, §4.1, with maximum
instead of summation to select a single path. This is forced alignment, not CTC
prefix decoding or the marginal probability of a target.
An exact predecessor tie prefers staying in the state, then moving one state, then moving two. An exact terminal tie prefers the final token to trailing blank. Accumulation uses double precision; emissions are float32. Negative infinity is an impossible emission, while NaN, positive infinity and positive log values are rejected. The caller supplies normalized log probabilities; the trellis does not normalize them. Traceback is bounded to 16 million cells, with two score rows. Cancellation is checked before validation and once per emission/trellis/traceback frame.
The result explicitly distinguishes failure from an empty target. Failure has negative-infinite score and no intervals; no prefix or guessed timing is returned. An empty target is the all-blank path, including a successful score-zero path for empty audio. Each successful nonempty target returns one inclusive-start, exclusive-end frame interval per target position. Repeated IDs remain distinct. Only frames occupied by that token count; intervening blank frames are not attributed to either neighbor. The score is the best path's log probability, not word confidence. Model adapters must establish frame-to-audio conversion and word boundary behavior separately; token occupancy is not calibrated word timing.
omniASR-CTC-300M adapter
The fp32 adapter loads the pinned publisher checkpoint as inert tensor data and
converts its 7-layer convolutional front end, 24-layer pre-normalized Transformer
and linear 9,812-character CTC head to OneModel format. It accepts mono 16 kHz
waveforms. The graph subtracts utterance mean and divides by the square root of
population variance plus 1e-5, then uses the checkpoint's convolution and layer
normalization parameters. The feature extractor has a 400-sample receptive field
and a 320-sample stride. The even-width positional convolution is trimmed by one
frame on the right. No oracle-exported graph or executable checkpoint object is
loaded.
The model vocabulary contains unique single-character pieces, an identity normalizer, no dummy prefix and no escaped whitespace. Written words are kept separate from the spoken target: casing and punctuation are retained in the result, while the alignment target is NFC-normalized, lowercased and excludes punctuation other than apostrophes. Combining marks are retained, typographic apostrophes become ASCII apostrophes, and hyphens become spaces. English integers from zero through 999,999,999 are expanded to cardinals internally (including comma-separated integers). This is not a general inverse-text-normalization engine: currency, decimal, date, acronym and language-specific number readings are not resolved. Unknown characters prevent that written word from receiving a timing.
Word occupancy starts at the first target character's occupied frame and ends at
the last character's exclusive end frame, converted with the fixed 20 ms stride.
No blank gap is spread over words, and no hand-tuned latency offset is applied.
A constrained word whose geometric mean occupied-frame probability is below
0.01 is left untimed. This rejection heuristic is not a calibrated probability
of word correctness. Confidence is not manufactured; existing word confidence,
speaker and other non-timing fields are retained by the transcript overload.
A call is bounded to 60 seconds; longer input is refused with a typed capability
error and must be split with its corresponding written words. File input is
bounded while decoding, before a model download, and is downmixed to mono.
Speech.Align tries Qwen3-ForcedAligner first for its eleven declared languages,
then this model for remaining untimed words. Omitted language uses CTC unless the
supplied transcript declares a covered language. An explicit model ID selects only
that aligner. First use downloads and verifies the selected publisher files and
reproduces the catalog's converted hashes. The string overload splits on whitespace. The transcript overload uses
Words in order without splitting them again. Corresponding segment word lists
are projected when they have the same written-word partition. Written text and
word boundaries are never reconstructed from the CTC hypothesis.
VadTranscriptionPipeline.TimingAligners is an ordered list of borrowed
aligners. Native timestamps win. Compatible forced aligners are tried in order
(for example Qwen3 before CTC); failed words retain existing DTW as the final
fallback. Compatibility requires Capabilities.ForcedAlignment and a matching
declared language. Whole-file transcription uses native timings first, then
Qwen3-ForcedAligner for zh/en/yue/fr/de/it/ja/ko/pt/ru/es, then CTC, then existing
DTW. Models are loaded lazily after the chunk's language is known; CTC is not
loaded if Qwen times every word. Failed or out-of-window alignment cannot invent
a boundary. Unsupported-audio and capability failures advance to the next method;
cancellation, corrupt models and I/O errors propagate. Successful words retain
per-word timing provenance. A lower-priority method never overwrites a successful
higher-priority timing. Qwen accepts more than 200 samples through 300 seconds;
CTC accepts at most 60 seconds. File recognition chunks remain bounded to 30
seconds; public file alignment retains its 60-second input bound.
Encoded-stream twins (Speech.AlignAsync) leave the caller's stream open. They
perform asynchronous bounded I/O to a temporary seekable file, capped at 512 MiB
of encoded bytes, before synchronous PCM decoding/inference. The temporary file
is removed on completion, cancellation or failure. Unsupported numeric readings
(including decimals, currencies and leading-zero strings) are left untimed, not
silently assigned the timing of a different spoken number.