🎙 Transcript Studio

Transcribe any English video — any length — into SRT / TXT / VTT / JSON, with per-speaker labels you can rename (Speaker 1 → "Alice"). 100% local: no cloud, no API keys, no accounts.

Features

• 🗣 Transcription — faster-whisper (offline, free) with Silero VAD for clean timestamps. English by default, --language auto to detect. Multi-hour videos work; progress is shown live.
• 👥 Speaker detection — a real speaker-verification model (NeMo Titanet via sherpa-onnx, ONNX, no torch) embeds every transcript segment and clusters the voices. The ~40 MB model downloads automatically on first use (cached in %LOCALAPPDATA%\transcript-studio\models). If it can't be installed, the app falls back to lightweight mel-based voice profiles. Built for panels and interviews: auto-detects up to 6 speakers, or force an exact count up to 8 (host + 7 guests).
• ✏️ Rename speakers — labels come as Speaker 1/2/… ordered by first appearance; give them real names in the UI or with --rename and the names are baked into every exported file.
• 📦 Exports — SRT and TXT (with timestamps + speaker names), plus WebVTT and JSON as extras. Subtitle cues are auto-split to ≤6 s / ≤110 chars.
• 🌐 Input — local video or audio files, uploads, or any URL (via yt-dlp).
• ⚡ Caching — re-runs of the same file are instant (transcript + embeddings are cached; changing names/speaker-count re-exports immediately).

Install

Requirements: Python 3.11+, FFmpeg on PATH.

cd D:\Cline\transcript-studio

# isolated env (recommended — uv can fetch Python 3.11 itself)
uv venv .venv --python 3.11
uv pip install -r requirements.txt -p "D:\Cline\transcript-studio\.venv\Scripts\python.exe"
uv pip install -e . -p "D:\Cline\transcript-studio\.venv\Scripts\python.exe"   # adds the `transcript-studio` command

# or once published to PyPI:
# pip install transcript-studio

ImportError: DLL load failed ... onnxruntime_pybind11_state — this machine's System32 only has the 2015 MSVC runtime. Copy the newer set shipped with Python 3.13 next to the ONNX DLLs (no admin needed):

$src = "$env:LOCALAPPDATA\Programs\Python\Python313"
$dst = "D:\Cline\transcript-studio\.venv\Lib\site-packages\onnxruntime\capi"
Copy-Item "$src\msvcp140*.dll","$src\vcruntime140*.dll","$src\concrt140.dll" $dst

Quick start

# UI (opens http://localhost:8502) — double-click the desktop shortcut
python -m transcript_studio ui

# CLI: SRT + TXT next to the video, auto speaker count
python -m transcript_studio run "C:\Videos\podcast.mp4"

# rename speakers, pick formats, force the speaker count
python -m transcript_studio run "C:\Videos\podcast.mp4" `
    --rename "Speaker 1=Alice,Speaker 2=Bob" `
    --speakers 2 --formats srt,txt,vtt,json -o D:\outputs

# long video: auto-parallel (up to 4 workers) is already the default
python -m transcript_studio run "D:\Videos\podcast-ep12.mp4"

# fastest possible local run (coarse cue timing) — combine batching + workers
python -m transcript_studio run "D:\Videos\podcast-ep12.mp4" --fast --workers 4

# trickier audio: bigger model (slower on CPU), English-only
python -m transcript_studio run "D:\Videos\noisy-panel.mp4" --model small.en --speakers 4

# from YouTube (audio-only download, fast)
python -m transcript_studio run "https://youtu.be/XXXX"

# options
python -m transcript_studio run --help

CLI flags (run)

Speed (measured on this machine)

Measured on an Intel i5-8350U (4 cores / 8 threads, no GPU) with the default base.en model — WER is word error rate against known ground truth:

What that means in practice:

• Short clips (up to ~8 min) run single-process — process overhead isn't worth it.
• Long videos automatically use 4 parallel workers (--workers 0 = auto): chunks are cut at the quietest point near the boundary, each chunk decodes 6 s of overlap on both sides, and only segments whose midpoint falls inside its keep-zone are merged back — so the result has neither gaps nor duplicates. Measured: ~3× faster and slightly more accurate than a single pass (0.97 % vs 1.36 % WER).
• --fast adds batched decoding (≈40 % faster on short audio; 16.9× combined with workers on long audio), but it produces longer subtitle cues whose inner timing is interpolated — good for a quick text transcript, not for precise subtitles (its text accuracy also drops a little: 3.2 % vs 1.0 % WER).
• Close other heavy apps first. On this 4-core CPU, background processes (Streamlit servers, browsers, agents) can cut throughput by 2× or more — during testing one orphaned server alone was burning 2 full cores.
• Big models (distil-large-v3, large-v3-turbo, large-v3) need a GPU to be fast — measured: distil-large-v3 runs at 0.85× realtime on this CPU (slower than real time), while base.en runs at 6×.

If you need real speed: a cloud API key

This CPU has a hard ceiling of roughly 5–17× realtime. A hosted speech API removes it entirely — e.g. Groq's whisper-large-v3-turbo (very fast, ~$0.04 per hour of audio) or OpenAI's gpt-4o-transcribe-diarize (fast and speaker-labelled in one call). There is no "Cline API" to call — Cline is this coding agent, not a speech service. If you drop a key into the environment (GROQ_API_KEY / OPENAI_API_KEY), I can wire a --backend groq|openai path into the app in a few minutes.

Tests

python -m pytest        # 27 tests: exporters, clustering, embeddings, sources

tests/make_two_voice.py and tests/make_multi_voice.py synthesize 2- and 4-voice conversations with ground-truth turn timelines (edge-tts, needs network); tests/eval_diarization.py scores the real pipeline against them — see the panel table below.

Multi-speaker panels (3–4+ guests)

Works out of the box — leave Speakers = Auto, or pass the exact count when you know it (--speakers 4, UI dropdown 1–8) to be extra safe:

python -m transcript_studio run "C:\Videos\panel.mp4" --speakers 4 --rename "1=Host,2=Ann,3=Ben,4=Chloe"

Measured on synthetic ground truth (same pipeline as production):

Reproduce it yourself:

python tests/make_multi_voice.py        # needs network (edge-tts) — builds a 4-voice clip
python tests/eval_diarization.py --audio tests/data/multi_voice.wav `
    --truth tests/data/multi_voice.truth.json --speakers 4 --diagnose

Guidance: prefer Auto or the true count. Forcing a higher number than reality doesn't mix people up — it just fragments one voice into two labels (e.g. 8 labels for 4 people, each still internally consistent).

How speaker detection works (and its limits)

1. Whisper segments are extended into the half-gaps around them (silence only — never a neighbour's speech), chunked to ≤30 s windows, and embedded by Titanet into 192-dim L2-normalized vectors.
2. Vectors are clustered with k-means (deterministic); the speaker count comes from silhouette scoring with a minimum-cluster guard, or is forced by --speakers.
3. Labels are ordered by first appearance → Speaker 1, Speaker 2, …

Embeddings are content-invariant, so short punchy lines from the same voice stay together (a plain pitch/timbre heuristic could not tell that apart). Honest limits: overlapping speech (several people talking at once) is assigned to the dominant voice of that segment; Whisper may merge a very short backchannel ("Agreed.") into the next turn — such segments go to the majority speaker. Voices that are literally identical (clones) may merge; the fallback mel profile is cruder — install sherpa-onnx (it's in requirements.txt) for the good path.

Use it from another computer (MacBook / phone on the same Wi-Fi)

The UI is a local web server, so any device on your network can drive it — the heavy work still runs on this laptop.

1. Start the server on the laptop — double-click Start_Transcript_Studio.bat (binds 0.0.0.0:8502, so it accepts connections from other devices).
2. Allow it through the firewall — once — double-click Allow_LAN_access.bat and accept the UAC prompt (adds an inbound TCP-8502 rule; needs administrator, which is why it's a separate file).
3. On the MacBook, open http://192.168.0.203:8502 (check the laptop's current Wi-Fi address with ipconfig; it changes if the network does). Upload the audio/video → pick formats → download the .srt it produces.

Server settings live in .streamlit/config.toml (address = "0.0.0.0", port = 8502, maxUploadSize = 4096 MB).

Away from home: install Tailscale on both machines and sign in — the MacBook then reaches the laptop from anywhere at http://<laptop-tailscale-name>:8502, still private to your own devices, and no firewall rule or public URL is involved.

License

MIT — see LICENSE.