Skip to content

Latest commit

 

History

History
132 lines (111 loc) · 5.2 KB

RELEASE.md

File metadata and controls

132 lines (111 loc) · 5.2 KB
Raw

1.3.0

  • Add input option "language" that can be passed at each request
  • Add result of language detection (or given language) in the output, for each segment
  • Add speaker identification ("speakerIdentification" option in "diarizationConfig")

1.2.12

  • Do not fail when asking to convert numbers with env. variable LANGUAGE=*

1.2.11

  • Improve heuristics to merge transcription and diarization results (for words in between two speaker turns)

1.2.10

  • Add heuristics to avoid too long speech segment sent to STT (limit risk of memory overflow)
  • Fix failure with token "- Et"

1.2.9

  • Avoid a 1H timeout that was causing celery task to re-run (and fail)
  • VAD: Improve heuristics about audio segment durations to better adapt to Whisper setting (minDuration=30)
  • Preserve exponents ("²") in word normalization

1.2.8

  • In full transcription: proper normalization of spaces before/after traditional punctuation marks (for French and English at least)
  • In word normalization: improve distinction of characters (word / punctuation / symbol that can be pronounced / garbage symbol)
  • Fix typo fr_FR -> fr-FR
  • Fix inconsistency in transcription confidence score (now always computed from word confidence scores, not segment confidence scores)

1.2.7

  • Remove punctuations in words (to avoid spaces as in "allez-vous ?")

1.2.6

  • Fix possible worker conflict when multiple workers are running on the same file (the audio file could be deleted by the worker that first finishes)
  • Fix speaker segment splitted in two when diarization detects another speaker with no word assigned.
  • Fix a bug in the formatting progression "status" ("StepState.PENDING" -> "pending"), which was introduced because python version was not fixed in Dockerfile (python 3.11 changes behaviour when converting enum to string)

1.2.5

  • Add options for VAD (minimum duration of segments, ...)

1.2.4

  • Fix corner case of empty transcriptions
  • Fix corner cases to assign words to speaker turns (overlapping diarization segments, words in between two segments)

1.2.3

  • Added multifiles route and processing.
  • Changed straddling word diarization resolve
  • Added diarization results within transcription result.

1.2.2

  • Added recover to redis search index drop.

1.2.1

  • Added Bearer Authentication to swagger.
  • Updated README.

1.2.0

  • Added timestamp interpolation for non-consecutive diarization segments.
  • Added Makefile for styling
  • Refactored code to PEP8 (black)
  • Reorganized repository folder structure.
  • Added service discovery for subtasks
  • Added service resolve and service resolve policy
  • Added task logs and log query route
  • Added possibility to upload a timestamps file.

1.1.2h1

  • Fixed convertnumber converting spk id 1
  • Fixed usersub not applied to subtitles
  • Fixed text cleaning and substitutions not applied to chunks of subtitle.

1.1.2

  • Added raw_return and convert_number to VTT and SRT format
  • Removed accept header check on /job/ route
  • Cleanup

1.1.1

  • Added: Text normalisation.
  • Added: Text to Number.
  • Added: Result presentation options as query string.
  • Added: MongoDB error handler.
  • Changed: Steps progression.
  • Updated: README
  • Updated: API specs.
  • Updated: transcription_request test script.

1.1.0

  • Added: A new route has been added /results/{result_id} allows to fetch transcription result and to specify the result format.
  • Changed: MongoDB server availibility timeout check greatly reduced to prevent hanging when mongo is unavailable.
  • Changed: The /job/{job_id} route now returns a ressource_id to be fetch on the /results/{result_id} when the task is completed.
  • Changed: Diarization is ignored when number of speaker is 1
  • Changed: GUNICORN_WORKER replaced with CONCURRENCY.
  • Fixed: Transcription worker concurrency is now set using CONCURRENCY env variable.
  • Updated: README.
  • Updated: Swagger's document.
  • Removed: no_cache request option has been removed.

1.0.3

  • Added: Subtitling return format for VTT and SRT
  • Added: Accept headers for subtitle formats
  • Added: jobid in result database
  • Changed: segment in TranscriptionResult will be equals to raw_segment in absence of postprocessing
  • Added: fetch result in db using jobid
  • Moved: transcription related file to workers/utils
  • Updated: README
  • Removed: no longer used formating.py file
  • Removed: SubtitleConfig in TranscriptionConfig

1.0.2

  • Added force_sync param for forced synchronous call
  • Added vad processing to split large files into subfiles
  • Added password variable for the service broker
  • Changed API to the TranscriptionConfig format.
  • Changed results return format
  • Updated test_transcription.py
  • Fixed wavefile not being converted when samplerate was wrong
  • Removed flower
  • Updated swagger to OpenAPI 3.0 and added new specifications.git

1.0.1

  • Added wait-for-it for service dependencies
  • Added LICENSE
  • Added README
  • Added swagger
  • Fixed post-processing failing with speaker diarization
  • Fixed transcription task initial state not returning proper format
  • Removed unecessary ENV variables
  • Moved test/ to repository root

1.0.0

  • Initial version
  • Allow client to perform asynchronous transcription request
  • Results are stored in a database