dinnote 0.2.0

Audio denoising, VAD, speaker diarization, and transcription pipeline using Demucs, Silero VAD, pyannote, and Whisper

dinnote audio transcription

Processes audio through a four-step pipeline to produce a transcription JSON with per-speaker diarization: denoising (Demucs), voice activity detection (Silero VAD), speaker diarization (pyannote), and transcription (Whisper).

Installation

pip install dinnote

This installs CPU-only torch by default. For GPU acceleration, install the CUDA build first:

pip install torch torchaudio --index-url https://download.pytorch.org/whl/cu124
pip install dinnote

dinnote uses CUDA automatically when available and warns at startup if it is not.

On first run, dinnote copies default config files to your platform config directory:

• Windows: %APPDATA%\dinnote\
• macOS: ~/Library/Application Support/dinnote/
• Linux: ~/.config/dinnote/

Edit config.yaml and vocab.txt to customize settings.

Speaker diarization requires a HuggingFace token with access to pyannote/speaker-diarization-3.1. Set it via diarize.hf_token in config.yaml.

CLI usage

dinnote input/audio.mp3            # single file
dinnote input/                     # all audio files in a folder
dinnote input/audio.mp3 -f         # force re-run all steps
dinnote input/audio.mp3 -c path/to/config.yaml   # custom config
dinnote input/audio.mp3 -o results/              # custom output dir

Each step checks whether its output already exists and skips it if so. Use -f to force all steps to re-run.

Output is written to output/<filename>/ and contains:

• <filename>_denoised.wav (vocals isolated from background noise)
• <filename>_vad.json (detected speech segment boundaries)
• <filename>_diarization.json (per-speaker turn boundaries from pyannote)
• <filename>_transcription.json (final transcription with timestamps and speaker labels)

Python API

from pathlib import Path
import dinnote
from dinnote import PipelineConfig, VadConfig, DiarizeConfig, TranscribeConfig

# Run the full pipeline with defaults
dinnote.process_file(
    input_path=Path("recording.wav"),
    output_dir=Path("output"),
)

# Custom config
config = PipelineConfig(
    vad=VadConfig(threshold=0.4, max_segment_length_sec=20),
    diarize=DiarizeConfig(num_speakers=2),
    transcribe=TranscribeConfig(model="small", language="en"),
)
dinnote.process_file(Path("recording.wav"), Path("output"), config=config)

# Or use individual stages
from dinnote import denoise, vad, diarize, transcribe

denoised      = denoise.run(Path("recording.wav"), Path("output/recording"), config={})
vad_file      = vad.run(denoised, Path("output/recording"), config={})
diarization   = diarize.run(denoised, Path("output/recording"), config={})
result        = transcribe.run(denoised, Path("output/recording"), config={}, diarization_path=diarization)

Configuration

denoise:
  model: htdemucs        # htdemucs | htdemucs_ft | mdx | mdx_extra | htdemucs_6s

vad:
  threshold: 0.5         # 0.0–1.0, higher = requires clearer speech
  min_speech_duration_ms: 250
  min_silence_duration_ms: 100
  padding_ms: 500
  max_segment_length_sec: 30
  merge_within_sec: 1.0

diarize:
  # hf_token: hf_...
  num_speakers: null     # fix speaker count or leave null to let pyannote estimate
  min_speakers: null
  max_speakers: null
  min_turn_ms: 200       # turns shorter than this are discarded (ms)

transcribe:
  model: base            # tiny | base | small | medium | large
  language: en           # set to null to auto-detect
  temperature: null      # null = Whisper fallback sequence, 0 = greedy
  no_speech_threshold: 0.6
  logprob_threshold: -1.0
  compression_ratio_threshold: 2.4
  condition_on_previous_text: false
  vocab_file: null       # path to domain-specific vocabulary, defaults to vocab.txt in config dir

Add domain-specific vocabulary to vocab.txt to improve transcription accuracy on unusual words and jargon. For noisy or technical audio, set temperature: 0 to disable Whisper's fallback to higher-temperature decoding, and consider filtering out common hallucinations specific to your dataset.

If num_speakers is known in advance, setting it gives more reliable diarization. Otherwise use min_speakers/max_speakers to constrain the range, or leave both null to let pyannote estimate freely.