uh!ive MRCP releases notes

1.22.0 — 2022-09-XX

1.21.0 — 2022-09-08

  • Builtin Date: recognize more expressions like "samedi dans 5 jours", "lundi de la semaine prochaine", "mercredi dans deux semaines", etc…
  • Builtin Boolean: give more priority to the strongs, dedicated "yes" word in case of ambiguity like "pas de soucis oui"
  • Basic support for Spanish (transcribe, text2num, spelling/*, boolean)

1.20.0 — 2022-08-08

1.19.0 — 2022-07-05

  • Spelling grammars: fixed confusion of I as Y.
  • Fixed handling of apostrophe in the postal_address grammar

1.18.0 – 2022-06-13

  • New parameter for speech/spelling/*: sep to specify the letter/number separator (or none) when no regex or length are given.
  • Bug fixes:
    • speech/spelling/mixed_with_punct now correctly recognizes and renders espace
    • speech/spelling/* ASR transcript y'a is now interpreted as i a
    • speech/date fixed ASR mis-transcriptions of month names mai and juin

1.17.0 – 2022-04-28

  • New and better Voice Activity Detection.

1.16.0 – 2022-04-11

Features:

  • More confusions resolved in spelling grammar.
  • Punctuation is not enabled by default on legacy grammars anymore (it caused regressions).
  • A new spelling grammar with support for punctuation has been added: builtin:speech/spelling/mixed_with_punct.
  • A new regex operator: ||; it's like | but it searches till the end of the utterance and it properly enforces precedence of the left-most alternatives.

Fixes:

  • Set error status on RECOGNITION-COMPLETE when an error occurs.

1.15.0 – 2022-02-07

  • New grammar: builtin:speech/date, fully functional in French, and partially in European English (numeric dates only in that case).

1.14.0 — 2022-01-26

  • Fixed EN support regression
  • Support for punctuation and diacritics in spelling grammars
  • More robust START-OF-INPUT

1.13.0 — 2022-01-10

  • less sensitive VAD
  • text2num fix of "lost" numbers after a silence
  • support for custom language extensions

v1.12.4 — 2021-12-08

  • better recognition in French spelling for J, G, T, Q, K and 100

v1.12.2 — 2021-11-22

  • Fixed Unsigned arithmetic overflow in timers that caused the recognition to hang when input timers are not started and the match is short.

v1.12.1 — 2021-10-25

  • Faster no-match conclusion thanks to our Speech-Nomatch-Timeout proprietary extension
  • Improved filtering of noise and low confidence words
  • finer heuristics to eliminate false positives at beginning of spelled things
  • no more no-match or partial-match returned by hotword (RFC)
  • sacrifice a little latency for better accuracy

v1.11.7 — 2021-10-08

  • Fixed keyword and boolean grammars to avoid premature rejection of speech

v1.11.5 — 2021-10-04

  • Deferred start-of-input
  • Fixed boolean serialization

1.11.4 — 2021-10-01

Fixes:

  • 'Recognition-Mode' must have priority in all cases
  • In the postal address builtin, fixed parts of the city leaking into the street
  • Fixed VAD for very low energy signals
  • Fixed regression in session handling preventing more than one RECOGNIZE in a session
  • Fixed a bug in GET-PARAMS that was returning the wrong value for confidence_threshold
  • SET-PARAMS correctly sets default parameters values now, and RECOGNIZE change to those value are not persisted beyond

Improvements:

  • better acuracy in most cases for spelling builtins by disabling rescoring
  • new buitin postal_address deprecates previous address builtin
  • Internal refactoring and optimizations
  • Support legacy hotword parameters in Vendor-Specific-Parameters header

v1.6.1 — 2021-06-07

  • New Voice Activity Detection algorithm for better accuracy and robustness to background noise (Sensity-Level is now ignored).
  • Improved spelling builtins to distinguish homophones through context (e.g. "de" vs. "deux").
  • Improved builtin address with structured interpretation that is now the preferred way to match addresses in production.

v1.5.4 — 2021-04-12

v1.5.3 — 2021-03-22

  • better session handling
  • improved builtin Address (more addresses are recognized)
  • fixed filtering of ASR tags

v1.5.0 — 2021-02-08

Features

Fixes

  • fixed parasite accented characters in spelling builtins
  • fixed letter 'x' transcribed as "ex" in spelling builtins
  • fixed premature no-match in normal mode

v1.4.2 — 2021-01-15

  • Properly close session after early no match
  • Typo fix in nomatch respond body that prevented correct parsing.

v1.4.0 — 2021-01-12

v1.3.0 — 2020-12-22

  • Normal mode and hotword mode full support according to RFC
  • Stricter RFC compliance of timers for Recognition-timeout, hotword-[min|max]_duration
  • Support for statuses success-maxtime, hotword-maxtime, partial-match-maxtime
  • Provide interpretation score in ASR score on <instance> and <input> respectively of output payload:
  • Now return appropriate body for no-match and no-input + scores interpretation & ASR.
  • Support for length parameter on spelling builtins.
  • Support for ZipCode builting

v1.2.0 — 2020-09-16

  • Bug fix: stream silence while waiting for start of input
  • Support for:
    • Hotword-Min-Duration (RECOGNIZE method only)
    • Hotword-Max-Duration (RECOGNIZE method only)
    • Recognition-Mode (RECOGNIZE method only)
  • No partial match for spelling grammars

v1.1.0 — 2020-09-10

  • Legacy mode builtin:speech/text2num to restore behavior of pre-0.13 versions.
  • Improved builtin:spelling/letters to handle double letters said as "2 + letter".

v1.0.0 — 2020-08-24

  • English support for plain ASR
  • Builtin ASR model configuration
  • English support for Spelling builtin
  • Log errors with appropriate log priority
  • Heartbeat to avoid TCP timeout between RECOGNIZE commands

v0.13.0 — 2020-08-11

  • New backend (Voxo-broker), for better scalability and lower latency.
  • 60ms bufferization (for lower load)
  • New spelling builtin grammar:
    • more robust,
    • with support for decimal numbers.
  • Removed obsolete configuration flag asr_always_matches.