speech-splitter

Turn any audio (or video) into a TTS training dataset: drop files in, press Start, get clean sentence-length clips with matching transcripts and a Hugging Face-ready metadata.csv.

What it does

1. Transcribes with faster-whisper (CTranslate2) using word-level timestamps — any language, CPU or NVIDIA GPU, models download on first use.
2. Splits by sentence (. ! ? in any script, numbers like 3.5 never split). Short sentences are merged with the next one, overlong ones are split at the longest pause; clips always end at the last word, never at the next word's start.
3. Filters with YAMNet (AudioSet classifier, runs on onnxruntime): music, singing, speech over music, noise, silence — plus Whisper's own word confidence and optional drop first/last segments for intros / outros.
4. Exports each clip with boundary buffers that extend into the surrounding pause but never into the neighbouring word, silence trimming, stray-fragment removal and optional cutoff detection (clips with a clipped first/last word are skipped or tagged _CUTSTART / _CUTEND).

output/
  audio/        speaker_0001.wav …        (wav / flac / mp3, original or chosen sample rate)
  transcripts/  speaker_0001.txt …
  metadata.csv  file_name,transcription,duration   <- Hugging Face AudioFolder
  metadata.json full per-clip info (confidence, classification scores, flags)
  manifest.txt  audio|transcript pairs
  logs/

Install

Grab a prebuilt binary from the latest release (Windows / Linux / macOS, no Python needed), or run from source:

pip install -r requirements.txt
python speech_splitter_gui.py        # GUI   (run.cmd / run.sh do the same)
python speech_splitter.py --help     # CLI

You also need FFmpeg in PATH or next to the app for anything other than plain WAV input.

GPU (NVIDIA)

faster-whisper runs on CUDA 12 and needs cuBLAS 12 + cuDNN 9. Either pip install -r requirements-cuda.txt, or use Model → Download CUDA libraries in the app (official NVIDIA wheels from PyPI, ~1 GB, resumable, stored in a cuda folder next to the app). The Model tab shows the GPU, the driver and which libraries were found; leave Device on auto. Expect ~10–30× realtime with large-v3-turbo on a mid-range card vs ~1–3× on a CPU with small.

Settings

Tab
Model	model (any size or Hugging Face repo id), language / auto-detect, device, compute type, beam size, VAD, initial prompt, model folder, transcription cache, YAMNet + CUDA downloads, GPU status
Segmentation	min / max clip length, merge short sentences, what to do with overlong ones (split / tag / skip), start & end buffers, cutoff detection (mode, sensitivity, skip / tag / keep)
Quality	min word confidence, music / mixed / noise / silence thresholds, edge-segment removal
Export	output folder, speaker id (numbering continues across runs), wav / flac / mp3, sample rate, silence trimming, fragment removal

Presets menu: Fast / Balanced / Strict, plus save / load / import / export your own and save as defaults.

Transcriptions are cached next to the source as <file>.transcription.json (tagged with the model used) — re-running with different segmentation / filter settings is instant.

CLI

python speech_splitter.py audiobook/ -o dataset --model large-v3 --language cs --speaker-id jan \
    --min-length 1.5 --max-length 12 --check-cutoff --filter-noise --min-confidence 0.6

All GUI options exist as flags (--help). Folders are searched recursively.

Tips

• Words still clipped? Lower --cutoff-sensitivity, or raise the buffers a little (they can't cut into words anyway).
• Too many clips rejected? Raise --music-threshold / --mixed-threshold, lower --min-confidence, or turn off --filter-noise for room-tone recordings.
• Proper names wrong? Put them into the initial prompt; Whisper copies their spelling.
• Garbage at file edges (jingles, "thanks for listening")? Drop first / last segments.
• The tools/ folder has helpers for chunking large datasets into parquet shards and converting wav → mp3.

Build

pip install -r requirements.txt pyinstaller && python build.py produces dist/speech-splitter-<version>-<os>-<arch>. The GitHub workflow builds all three platforms on every tag and attaches them to the release; the Linux build runs a headless self-test (SPEECH_SPLITTER_SELFTEST=<audio file>).

Changes in 3.0.1

• Helper processes (ffmpeg / ffprobe decoders started by pydub, nvidia-smi) are tracked and killed when the app quits, is stopped, or crashes — Windows Job object, Linux parent-death signal
• No partial output files: clips, transcripts, metadata.csv / metadata.json / manifest.txt, the transcription cache and presets are written to a temporary name and renamed into place when complete; the run log is .log.part until the run finishes
• Adding files or folders never queues the same path twice (status bar reports "n already in queue")
• Folder scans, the queue, output listings and presets use natural order (take2 < take10)

Changes in 3.0

• faster-whisper replaces WhisperX (no torch / pyannote; word timestamps built in; CUDA libraries downloadable in-app)
• YAMNet runs on onnxruntime instead of TensorFlow — the app is a ~270 MB standalone download instead of a Python 3.10 virtualenv with 5 GB of dependencies
• new GUI in the style of whisperer / videer: queue with per-file status, live progress, presets, GPU panel, model downloads
• sentence splitting works for all languages (was ASCII-only); short sentences are merged instead of dropped; long sentences are split at pauses instead of just being tagged _LONG
• end buffers no longer cut into the next word; metadata.csv is now valid for Hugging Face AudioFolder (audio/… paths, transcription column); transcription cache is invalidated when the model changes
• fixed: the CLI could not start (duplicate --check-cutoff / --cutoff-action flags, missing --min-confidence), crash when YAMNet was unavailable, filter silence and trim silence toggles being ignored
• DeepFilterNet enhancement was removed (needed torch and a CPU-only monkeypatch); run your audio through a denoiser first if you need it