v2.0.19

vachan-access v2.0.19

A . New APIS

Forgot Password (Recovery by Code) — 3 APIs

1. Initiate recovery
   POST /v2/auth/user/forgot-password
   Triggers a Kratos recovery flow (method=code) and emails a 6-digit code to the user.

2. Verify recovery code
   POST /v2/auth/user/verify-recovery-code
   Validates the code and returns the settings_flow_id (server-side step to proceed with password change).

3. Set new password
   POST /v2/auth/user/reset-password
   Completes the flow by posting the new password to Kratos settings API .

B. Enhancement

Simplified password rules in kratos.yml:

1. Users can choose very short passwords (6+ characters)
2. Kratos will not check passwords against the public “Have I Been Pwned” breach database.
3. Kratos will allow passwords similar to the username/email

v2.0.18

vachan-ai 25.10.16-beta

Text-to-Speech:

• Added new model Indic Parler-TTS
  • Supproted languages: Assamese, Bengali, Bodo, Dogri, English, Gujarati, Hindi, Kannada, Konkani, Maithili, Malayalam, Manipuri, Marathi, Nepali, Odia, Sanskrit, Santali, Sindhi, Tamil, Telugu, and Urdu.
  • 'language' field is not required for indic-parler-tts model.
  • Speaker Voice: Generated audios use the model's default speaker voices. Can be changed through description. For more details refer to link .
  • New input field 'description': A detailed description of how the speech should sound, e.g., "Leela speaks in a high-pitched, fast-paced, and cheerful tone, full of energy and happiness. The recording is very high quality with no background noise." This field is relevant only for indic-parler-tts model. Can be left empty.
• Added finetuned Chatterbox model
  • Supproted languages: English, Hindi, Telugu and Malayalam
  • Voice Customization: Audios can be generated in desired voice by passing reference audio

Text translation :

• The output of document translation can be downloaded either in txt, docx or html format
• Now finetuned NLLB models can be used via APIs
  • nllb-english-zeme (one-direction)
  • nllb-english-nagamese (bi-directional)
  • nllb-gujrathi-koli_kachchi (one-direction)
  • nllb-hindi-surjapuri (bi-directional)
  • nllb-gujrathi-kukna (bi-directional)
  • nllb-gujrathi-kutchi (bi-directional)

Languages :

• Included Post Language API
• Added new language column ISO_639_1_code in the language table using alembic migration

Enhancements :

• Model dependencies are now loaded from an external configuration file, allowing deployment servers to modify them without changing the codebase.
• Mlflow upgradation
• Enhanced GPU prediction process
• Integrated Pyright type checks across the codebase and in the CI pipeline

Docs:

• Updated architecture diagram and documentation

Next release plan:

• Speech-to-Text : Optionally generate SRT/VTT files of the uploaded audio
• Speech-to-Speech : Cascaded pipeline using STT, MT and TTS APIs
• Text-to-speech: Chatterbox model with more language support
• Python version update

v2.0.17

Vachan-AI

New features

• TTS: Integrated orpheous wrapper
• STT: punctuated transcripts using LLM Api
• Voice clone: Integrated chatterbox wrapper

Enhancement

• Serving mode for VACHAN_AI_ENV to suppress warning logs during app startup
• Changed predict function param to model_input based on the mlflow warning that appear during the app start
• Orpheous model handles with or without reference audio
• Package updation-SQLalchemy,pydantic ,uvicorn ,pitz ,librosa,python-multipart,fastapi,pytest,pytest-mock,pylint,redis,rq,numpy,pydantic-settings,skypilot,llama-index,docx2txt,PyMuPDF,orjson
• Router modularization -seperate ai_apis into seperate files ( assets_apis.py, cloud_operations_apis.py, inference_apis.py, jobs_apis.py, languages_apis.py, model_apis.py)
• Added type hints for predict function in all wrappers ,for removing warnings from app start
• MT Cloud finetuning - Included training script and config file for MT cloud fine tuning
• Updated pylint version to v3.3.7
• Changed the input parameter name in voice clone API to input_transcripts

Docs

• Converted i/p params json to valid format in models csv
• Modified architecture documentation
• Updated app folder structure
• Container diagram, component diagram : included ray, modified training pipeline
• Moved auth csvs into a new folder within data
• Updated model csv and model documentation (Added chatterbox details)

v2.0.16

vachan-ai 25.4.22-beta

New features:

• API for audio enhancement using resemble-enhance
• Model logging feature for resemble-enhance
• New API for TTS
  • Accepts Reference audio
  • Accepts text as document(txt file), splits text into sentences, generate audios for sentences
• Text-to-speech with ToucanTTS (pre-trained and fine-tuned models)
  • Audios can be generated based on a reference audio
• Model logging feature for ToucanTTS
• Priority queue: Jobs can be handled based on assigned priorities

Enhancements:

• Audios can be generated in ogg format
• Enhanced audio generation
  • Enhancements using resemble-enhance
  • Sampling rate : 48khz
  • Bit depth : PCM_24
• Updated languages csv to include more languages(as toucan model supports 7000+ languages)
• 'model_name' field of all apis is a drop down now, listing the available models
• Detailed logs added

Code refactoring:

• Split inference core (having core functions of different services) into small modules for following features
  • Noise removal
  • Forced alignment
  • Audio segmentation
  • Speech-to-text
• Moved model specific env variables into a data class
• Included config file for the project
  • Moved frequently changing env variables into the config file
  • App can be configured with a config file that sits independent of the code base

v2.0.15

vachan-ai

• Parallel processing for handling multiple jobs
• Audios can be generated in mp3 format
• Model logging : model_uri is optional now
• Minor fixes
• Calendar versioning for the app

v2.0.14

vachan-ai

New features

• Device option added: APIS can be run in GPU and CPU @Jayasankar-kk
• API for retrieving served models details @Jayasankar-kk
• Audio splitting utils @anjalyv
• Model logging support for LLMs @Jayasankar-kk
• API for forced alignment @Jayasankar-kk

Enhancements

• Output files of TTS, S2S, voice-clone and noise removal will have same name as that of input @NoelSudhish
• Improved segmentation accuracy by post-processing the segments info @anjalyv
• Package update @shaziya
• Other fixes @NoelSudhish , @shaziya

Documentation

• Detailed model documentation @shaziya, @anjalyv
• Updated swagger doc, readme @anjalyv @shaziya

v2.0.13

vachan-ai

Bug fixes:

• Resolved issue with mp3 audio files @anjalyv
• Raising error related to issues on the model serving side @Jayasankar-kk
• Updated model info record @shaziya
• MT: Enhanced pdf, txt file processing @anjalyv
• Minor fixes

v2.0.12

vachan-ai

• Model logging: Model wrapper for finetuned mms-tts models
• Jobs table: Included new columns User_id, Job_creation_time and Job_updation_time
• API for retrieving job history based on user_id
• Model serving using Rayserve
• MT: Text splitting utils using nltk
• MT: Document translation
• Audio segmententation api
• Included model metadata along with the model while logging

v2.0.11

What's Changed

• Updated the make file by @Joel-C-Johnson in #864
• Modified AWS ENV variables by @NoelSudhish in #872
• New release of vachan-ai [ 0.0.0-alpha.15] by @NoelSudhish in #874
  - API for background noise removal
  - Enhancement in logging models from s3 bucket
  - Model logging script for DeepFilterNet model (noise removal)
  - Fixed gitlab pipeline issue with Kratos by using docker profile

Full Changelog: v2.0.10...v2.0.11

v2.0.10

What's Changed

• Adding more details in deployment steps by @AthulyaMS in #862
• commented out the vachan-api-test container and Run test in workflow by @KetanKBaboo in #863
• auth based on flag for vachan-ai by @NoelSudhish in #866

New Contributors

• @KetanKBaboo made their first contribution in #863

Full Changelog: v2.0.9...v2.0.10