mobius-faster-whisper 1.1.0

This release is based on the original faster-whisper repo (release v1.0.3) and contains below additional features:

Turbo model support (#33)

• Based on openai/whisper#2363

Batching Support, Speed Boosts, and Quality Enhancements (Based on SYSTRAN#856)

• Batching support
• Faster feature extraction with torch STFT
• Quality Improvements:
  • Multi-segment language detection
  • Code-switching support
  • Consistency across runs
  • Reducing hallucinations

Bug Fixes

• fixes no_active_speech vad error (#36)
• limit torch version for cudnn8 (<2.4.1)(#33)

Latest release before merging to FW

Changes include removing the default FE method and replacing it with the torch-based method, restructuring components for ease of reliability, and reducing redundancy.

faster whisper v1.0.1 with mobiusml additions

Primer before faster-whisper PR:

1. Comply with CONTRIBUTING guidelines.
2. Added tests for batched transcription and multisegment language detection.
3. Added vad model by default for batched transcription.
4. minor fixes in the code/requirements.

faster whisper v0.10.0 with mobius additional capabilities

This version builds explicitly on faster_whisper 0.10.0 and has the following additional capabilities:

1. All Mobius features that were present in the previous release.
2. Support for batched inference (assuming Vad segments are fed as inputs) in streaming and batched output modes.
3. Support for multisegment language detection that is more accurate.

faster whisper v0.9.0 with additional features for mobius ASR v2.1

This release is based on latest faster-whisper project (v0.9.0), further changes include:

1. Adding multilingual support (Major)
2. Fixing seed for consistent results
3. Reduce hallucination by skipping ambiguous transcription segments
4. Adding numpy requirements

faster whisper 0.6.0 with multilingual capability, seed and fixes

Faster Whisper v0.6.0 with additional capabilities:

Multilingual support: Optional flag to support multilingual videos. The default output language is English. There is an option to set code-switched language as the output language.
Setting seed for ctranslate2 model: Useful for consistency reasons.
Skipping the segment if the avg_log_prob is too low: The current option also checks for no_speech_prob and ignores music/noise pieces.