Skip to main content
This is Hear’s cleanup layer: make a final transcript readable without slowing the live partials. Interim results are never rewritten. Every step is fail-open: a timeout, exception, or lexical-guard miss returns the previous text unchanged.
Hear sync/streaming supports the published language set. Number formatting and the English lexical cleanup flags are strongest on English. Unsupported cleanup operations remain conservative/no-op. Async jobs transcribe the same eight languages; the spoken language is auto-detected per call.

Flags

numerals is independent of the others. dictation, drop_fillers, and vocabulary are not implied by smart_format.

smart_format

Opt-in. Changes case and punctuation only. Word identity is guarded. If wrapping would add or drop a word, Hear returns the unformatted final.
List commas keep the spoken and. Run-on splits are conservative: they fire only with enough words before a new clause (the meeting/call/order/…, also / anyway / meanwhile, please send/call/…, or a late standalone i). Tag questions use a short closer (right, yeah, okay, isn't it, …). smart_format does not:
  • Invent emails, URLs, or street addresses
  • Expand or rewrite words (threepm stays threepm unless ITN already normalized it)
  • Remove fillers (that is drop_fillers)
  • Honor spoken “period” / “comma” (that is dictation)
  • Format non-English text

dictation

Separate from smart_format. Off by default. On English finals only, these spoken commands become punctuation:
  • period.
  • comma,
  • question mark?
  • new paragraph → a paragraph break
Turn smart_format on as well if you also want sentence capitalization.

drop_fillers

Off by default. Strips filled pauses (um, uh, umm, uhh, er) on English finals. Do not enable on legal or compliance audio by default. Those tokens can be evidence.

vocabulary

On a curated human evaluation set focused on named entities, vocabulary context raised entity recall from 70.9% to 89.3%. Aggregate word error rate on the same set rose from 23.5% to 24.9%. This is a measured entity-recall tradeoff, not evidence of a general accuracy improvement. Use a short list only when you know which distinctive terms are likely to appear.
Use vocabulary only when you know distinctive names, brands, products, or phrases that may appear in the audio. Streaming and async jobs use up to five effective terms. They trim whitespace, preserve the first spelling and order, deduplicate without regard to case, and ignore entries that are shorter than 4 characters, longer than 64 characters, longer than five words, or made only of common words. Vocabulary biases recognition toward your terms; it does not steer the output script (Hindi decodes render Devanagari regardless of how a term is spelled in the list). You can send vocabulary for one stream or job. You can also store an organization list with GET and PUT /v1/hear/vocabulary using the hear:configure scope. Stored terms remain inactive until enabled_for includes batch or hear_stream. When both are active, request terms come first. Stored suggestions fill any remaining slots up to five. The resulting list is fixed when the job is created or the stream opens. Changing organization settings does not change work that is already running.
Organization Hear vocabulary is not used by Omni. A managed Omni Agent can use its own explicit vocabulary list. Hear does not automatically import names from a CRM, dialer, or callee record.

Phone numbers

NANP grouping lives under numerals / ITN, not smart_format:
9-digit SSN-shaped spans and 16-digit card-shaped spans are never grouped.

Surfaces

Examples

On the stream, render partial greyed and only commit speech_final / final. That is how you get live captions that snap to punctuated text instead of flickering punctuation on every hypothesis. See Stream speech-to-text.

Failure behavior

Formatting is best-effort and fail-open. A timeout, exception, or lexical-guard miss returns the previous final unchanged. dictation and drop_fillers may change word count. When word timestamps cannot be realigned, the transcript still updates and the words array is cleared. Clients should not treat missing punctuation as an API error.

Next steps

Stream speech-to-text

Partials vs finals, commit, and endpointing.

Timestamped recording jobs

Word/segment offsets, subtitles, diarization, webhooks, and retention.

Conversation intelligence

Batch jobs, diarization, and post-call analytics.