v2.0.2

speechlib 2.0.2

Version 2 introduces major revamp of the API along some improvements to package quality and speed.

Architecture

• New Pipeline class replaces legacy Transcriptor API — instance-based, no provider kwargs at call time
• Abstract base classes (BaseDiarizer, BaseRecognizer, BaseASR) with full docstrings — plug in any backend
• All provider-specific settings live on component instances (PyAnnoteDiarizer, FasterWhisperASR, etc.)
• All legacy files deleted (speechlib.py, core_analysis.py, transcribe.py, speaker_recognition.py, wav_segmenter.py, etc.)

Performance

• Single-pass diarization — full file always diarized in one pass, preserving global speaker context
• Parallel transcription via ThreadPoolExecutor — segments transcribed concurrently across workers threads (CTranslate2/PyTorch release the GIL, so threads give real concurrency)
• BytesIO virtual filesystem — audio clips extracted into in-memory buffers instead of temp files on disk; no disk I/O, no cleanup, inherently thread-safe
• workers defaults to cpu_count - 1; pass 1 for sequential

ASR Backends as classes

• FasterWhisperASR — faster-whisper (CTranslate2), int8 quantization option
• WhisperASR — OpenAI Whisper
• CustomWhisperASR — local fine-tuned .pt checkpoint
• HuggingFaceASR — any HF ASR model; explicit torch_dtype per device (float32 on CPU, float16 on CUDA) to fix LayerNorm half-precision error; integer device index to fix meta tensor error
• AssemblyAIASR — cloud API; writes single temp file only when given a buffer (unavoidable)
  All backends accept str | BinaryIO in transcribe()

Pipeline customizability

• whole pipeline including diarize, ASR, speaker recognition stages are fully customizable
• use base classes BaseDiarizer, BaseRecognizer, BaseASR to implement your own solutions

Diarization

• PyAnnoteDiarizer — supports num_speakers (exact count) or min_speakers/max_speakers

Robustness

• Zero-duration and empty segments skipped before transcription (prevents reshape errors)
• Same guard in speaker recognition — empty clips skipped before SpeechBrain embedding extraction

Output

• output_format: "txt", "json", or "both"
• srt=True — writes SRT subtitle file alongside transcript
• Batch mode — pipeline.run([file1, file2, ...]) returns one result list per file

Build System is now Hatchling

• Hatchling provides a modern, fast, and PEP-compliant build system that replaces legacy setuptools with a fully declarative pyproject.toml approach, resulting in quicker, more reproducible builds, a cleaner project structure, and simplified packaging without the need for setup.py.

Contributors

@ThisisShashwat

v1.1.17

Minor improvements

v1.1.16

• fixed dependency issues
• using uv instead of pip

v1.1.12

just did a readme update :)

v1.1.11

• fixed numpy issues
• updated all dependencies to latest versions
• now using speechbrain 1.0 as Pyannote.audio 3.3.1 now supports it

v1.1.10

added flexible dependency versioning

v1.1.5

added missing cuda support for custom and hf models

release v1.1.4

fixed some issues in code

release v1.1.3

changelog:

1. added support to load custom whisper models from local disk
2. added support to use huggingface finetuned whisper models

release 1.1.2

1. added fp16 support for running normal whisper on gpu
2. updated readme