LocalTranscriber (.NET 10)

A small local voice-to-text starter that can:

• Record from a selected Windows capture device (WASAPI)
• Transcribe with Whisper via Whisper.net (GPU via CUDA runtime if available)
• Emit a Markdown file with structure
• Optionally polish the Markdown using a local LLM via Ollama

Prereqs

• Windows 10/11 (recording uses WASAPI)
• .NET 10 SDK installed
• If you want LLM formatting: Ollama running locally
  • Install Ollama and pull a model, e.g.: ollama pull llama3.2
• If you want cloud formatting via Hugging Face + Semantic Kernel:
  • Create a Hugging Face token and set HF_TOKEN
  • Example (PowerShell): $env:HF_TOKEN = "hf_xxx"

Quick start

From the repo root:

cd src\LocalTranscriber.Cli

dotnet restore

dotnet run -- devices

dotnet run -- record --device 0 --out ..\..\output\note.wav

dotnet run -- transcribe --in ..\..\output\note.wav --out ..\..\output\note.md --model SmallEn --format-provider auto

# Optional: use Semantic Kernel + Hugging Face instead of Ollama
dotnet run -- transcribe --in ..\..\output\note.wav --out ..\..\output\note.md --format-provider huggingface --hf-endpoint https://router.huggingface.co --hf-model openai/gpt-oss-20b:groq

# Optional: enable speaker labels quickly
dotnet run -- transcribe --in ..\..\output\note.wav --out ..\..\output\note.md --format-provider local --speakers 2

One-liner: record then transcribe

dotnet run -- record-and-transcribe --device 0 --wav ..\..\output\meeting.wav --out ..\..\output\meeting.md --model SmallEn

Blazor UI (record in browser + live progress)

cd src\LocalTranscriber.Web
dotnet run

Then open the app URL (for example http://localhost:5078) and use the recording panel.

Experimental browser-only mode (WebGPU)

• In the Blazor UI, enable Run In Browser (Experimental) when the capability badge reports support.
• Browser mode runs Whisper in-browser (Transformers.js) and can optionally run markdown cleanup via WebLLM.
• Browser mode requires a modern browser with WebGPU, AudioContext, and MediaRecorder.
• WebLLM/Transformers models are fetched by the browser at runtime, so first run can take time and requires internet access.

Notes

• Model download: the first transcription downloads the Whisper GGML model into:
  • %LOCALAPPDATA%\LocalTranscriber\models
• Accuracy vs speed:
  • SmallEn is a very good default.
  • MediumEn / LargeV3 are more accurate but heavier.
• Segment limit option:
  • Use --max-seg-length <int> to cap segment length (legacy --max-seg-seconds is still accepted).
• Convenience flags:
  • Use --format-provider auto|local|ollama|huggingface instead of juggling multiple boolean switches.
  • Use --speakers <int> as a shortcut for speaker labeling.
• Speaker labeling:
  • If a speaker count is not supplied by the caller, the app now auto-estimates speaker count and labels speaker swaps heuristically.
  • Speaker labeling is now tunable:
    • --speaker-sensitivity 0-100 (low = conservative, high = aggressive)
    • --speaker-min-score-gain
    • --speaker-max-switch-rate
    • --speaker-min-separation
    • --speaker-min-cluster-size
    • --speaker-max-auto
    • --speaker-global-variance-gate
    • --speaker-short-run-merge-seconds
  • In the Blazor Studio UI, these are exposed in Settings as sliders/advanced numeric controls.
• Formatting:
  • --format-provider auto tries Ollama first, then Hugging Face (if HF_TOKEN / --hf-api-key is configured), then local formatting.
  • --format-provider ollama or --format-provider huggingface force a provider, with local fallback if unavailable.
  • Formatting is tunable:
    • --format-sensitivity 0-100
    • --format-strict-transcript true|false
    • --format-overlap-threshold
    • --format-summary-min / --format-summary-max
    • --format-include-action-items true|false
    • --format-temperature / --format-max-tokens
    • --format-local-big-gap / --format-local-small-gap
  • In the Blazor Studio UI, formatter tuning is exposed via slider/toggles and advanced numeric fields.
  • If not, you still get a usable Markdown transcript via a deterministic local formatter.
• CLI output:
  • After transcription, the CLI also prints the transcript text to the terminal for quick verification.
• Web API caching (checksum/signature dedupe):
  • The Blazor server endpoint now computes a SHA-256 checksum for uploaded audio.
  • It also computes a request signature from audio checksum + normalized transcription/format/speaker settings.
  • Matching submissions are served from persisted cache instead of re-running normalization/transcription/formatting.
  • Cache entries are stored under src/LocalTranscriber.Web/output/cache.

Docker (Cross-Platform)

Run LocalTranscriber anywhere with Docker:

Quick Start

# Start web UI with Ollama
docker compose up web

# Or use GPU acceleration (requires NVIDIA Docker runtime)
docker compose --profile cuda up web-cuda

CLI Batch Processing

# Transcribe a file
docker compose run --rm cli transcribe --in /app/input/meeting.wav --out /app/output/meeting.md

# With GPU
docker compose --profile cli-cuda run --rm cli-cuda transcribe --in /app/input/meeting.wav --out /app/output/meeting.md

Available Images

Image	Description
local-transcriber:web	Web UI, CPU-only
local-transcriber:web-cuda	Web UI with CUDA GPU support
local-transcriber:cli	CLI, CPU-only
local-transcriber:cli-cuda	CLI with CUDA GPU support

Volumes

• ./models → /app/models — Whisper model cache
• ./output → /app/output — Transcription output
• ./input → /app/input — Input files (CLI only)

Building Manually

# CPU version
docker build --target web -t local-transcriber:web .

# CUDA version
docker build --target web --build-arg VARIANT=cuda -t local-transcriber:web-cuda .

Why Whisper.net?

Whisper.net is a .NET wrapper around whisper.cpp and supports multiple runtimes (CPU, CUDA, etc.) and a built-in GGML model downloader.

GitHub Readiness

This repo includes baseline GitHub project automation:

• CI build workflow: .github/workflows/ci.yml
• Dependency review on PRs: .github/workflows/dependency-review.yml
• PR auto-labeling: .github/workflows/pr-labeler.yml + .github/labeler.yml
• Label synchronization: .github/workflows/label-sync.yml + .github/labels.yml
• Release pipeline (tag-driven): .github/workflows/release.yml
• Dependabot updates: .github/dependabot.yml
• Issue and PR templates under .github/

Project governance/docs:

• CONTRIBUTING.md
• CODE_OF_CONDUCT.md
• SECURITY.md
• SUPPORT.md
• CHANGELOG.md
• LICENSE

First GitHub Setup Checklist

1. Update .github/ISSUE_TEMPLATE/config.yml discussion URL (<owner>/<repo> placeholder).
2. Update CODEOWNERS with real usernames/teams.
3. Push main, then run Label Sync workflow once to seed labels.
4. Create a tag (for example v0.1.0) to validate the release workflow.