LID: Terminology and adaptation

This article describes various ways of Language Identification adaptation.

Basic terminology

	Languageprint (*.lp file) – numeric representation of the audio, extracted from audio file for language identification purpose of (similar to “voiceprint”, but representing sound of the spoken language, not sound of the speaking person)
	Languageprint archive (*.lpa file) – multiple languageprints combined into single archive Creation of languageprint archives is not supported by SPE, these are supported as input only.
	Language model – digital characteristics of a specific language Language model can be trained from languageprints (*.lp), language prints archives (*.lpa), or from combination of both. NOTE about the word “model”: The LID language model should not be confused with LID technological model, like L4, L3, XL3, etc. which refer to the LID technology generation.
	Language pack – set of language models used for language identification

⚠ All these entities MUST use the same technological model!
For example, languageprints created using model L4 can be combined into languageprint archive and/or language model only with languageprints created using model L4… and language pack for model L4 must consist only from language models created using languageprints/archives of model L4.

 

Adaptation types overview

• Creating new language model from your own audio files, to add new language not supported out-of-the-box
  • at least 20 hours of audio is required, see requirements below
• Enhancing existing language model by adding your own audio files to existing built-in language
  • at least 5 hours of audio is required, see requirements below
• Creating custom language pack consisting of your chosen set of languages, both pre-trained or created from your audio files

Audio recordings requirements

• Format: WAV, FLAC, RAW with linear coding 16bit/8bit, sampling rate 8kHz+
• Wide variety of speakers (50+) of various age and gender is required, to ensure rich variety of “language sounds”
• Only single language in the dataset
  NOTE: mixing in a different language negatively affects the resulting recognition accuracy
• Audio length: ideally between 1 and 5 minutes of speech signal
  NOTE: it is not possible to train a language using just a few and long audio files (like 5 files, 1 hour each)
• Acoustic channels should be as close as possible to channel of intended deployment

---

Adaptation using REST API (SPE 3.38 or newer)

SPE 3.38 and newer include LID adaptation tasks in REST API, which makes the adaptation significantly easier than in previous versions.

Creating language model

Language model can be created from languageprints (*.lp) extracted from audio files, or from pre-trained language prints archives (*.lpa), or from combination of both.
Combination of both .lpa and .lp is used when enhancing existing language model – the .lpa is the existing language model and the .lps are created from your audio files.

• Extract languageprints from you audio files
  using GET /technologies/languageid/extractlp endpoint
• Create new (yet empty) language model
  using POST /technologies/languageid/languagemodels/{name} endpoint
• Upload languageprint- or languageprint archive file to the language model
  using POST /technologies/languageid/languagemodels/{name}/file endpoint
  • repeat this upload for all necessary files – e.g. when creating completely new language model from your own audio files, this would be hundreds or thousands of files (see audio requirements above)

More details are available in the REST API documentation: https://download.phonexia.com/docs/spe/#examples_languageid_create_lmodel

Creating language pack

• Get list of language models available for creation of your language pack
  using GET /technologies/languageid/languagemodels endpoint
• Create your custom language pack
  using POST /technologies/languageid/languagepacks/{name} endpoint

More details are available in the REST API documentation: https://download.phonexia.com/docs/spe/#examples_languageid_create_lpack