whispergram 0.8.2

Local, offline Telegram chat transcriber - voice & video via Whisper, screenshots via OCR, photos via a local vision model

whispergram

A Telegram chat — voice, video and photos — as one searchable transcript, fully local. Transcribe voice & round-video notes with Whisper (faster-whisper), read text from screenshots with OCR, and caption photo scenes with a local vision model — all merged into one chronological, LLM-ready file. 100% offline, no API key, no cloud.

Every line is tagged by sender and timestamp — voice, video and photos turned into readable text:

[2026-06-20 12:33] Alex (voice 14s): just finished the auth flow, take a look
[2026-06-20 12:35] You: nice, send the diagram
[2026-06-20 12:46] Alex (photo, described): a hand-drawn architecture diagram on a whiteboard | text: Login -> API -> DB
[2026-06-20 12:47] You (video-note 6s): looks great, let's ship it
[2026-06-20 12:47] Alex (sticker 👍)

Photos become text two ways: a local vision model captions the scene (on by default), and --ocr reads any text in the image.

---

Why?

An audio-heavy Telegram chat is unreadable and unsearchable — you cannot grep a voice note, and you cannot hand a folder of .ogg files to an LLM. The alternatives are worse: Telegram Premium transcribes one message at a time by hand, and cloud speech APIs upload your private audio to a third party. whispergram transcribes every voice and video note in one pass, entirely on your own machine, and weaves them back into the text timeline as a single file you can read, search, or feed to a model.

---

Features

Feature	Description
Voice + video notes	Both voice_message and round video_message notes are transcribed inline with the text
One merged file	A single merged_chat.md, chronological, every line tagged [time] sender
100% local & offline	faster-whisper runs on your machine — no upload, no API key, no account
Lossless mapping	Stickers, photos, animations, documents, music, locations, polls and contacts appear as markers — nothing content-bearing is dropped
Handles missing media	Notes excluded from the export are clearly marked [not exported], never fed to the model
All text shapes	Reconstructs plain, rich, and entity-based message text (links, mentions, custom emoji)
Dry-run	Preview the full merge with --dry-run — no model download, no GPU, instant
GPU or CPU	CUDA with automatic CPU fallback; a one-command Windows CUDA fix is built in
Auto-detect	Finds the export JSON (any filename) and the language per file
Regular videos	--video-files also transcribes ordinary video files' audio, not just round notes
Photo OCR	--ocr pulls text out of photos with local Tesseract — great for screenshots
Photo descriptions	Captioned automatically by the best installed local model — BLIP ([describe]) for photos, or Qwen2-VL ([describe-hq]) for photos + stickers + GIFs
Resumable	Progress is cached per file — close the terminal or crash, then re-run and it continues where it left off
Queue folders	Transcribe many exports in one command (models load once); --out-dir collects the results
Progress bar	Live done/total + ETA per folder
Round-trip verified	A rich synthetic export runs through the full pipeline and is diffed line-for-line; 85 offline tests on the Python 3.9–3.13 CI matrix

---

Quick Start

1. Install

Via PyPI (recommended):

pip install whispergram

Or clone for development:

git clone https://github.com/davidmalko87/whispergram.git
cd whispergram
pip install -r requirements.txt

You also need ffmpeg on your PATH:

# Linux:  sudo apt install ffmpeg
# macOS:  brew install ffmpeg
# Windows: choco install ffmpeg   (or: winget install Gyan.FFmpeg)

2. Export your chat from Telegram

Telegram Desktop → open the chat → ⋮ menu → Export chat history. In the dialog:

• Format: JSON (required — whispergram reads the JSON export, not the HTML one).
• Tick the media you want whispergram to use:

Export option	Tick it?	What whispergram does with it
Voice messages	✅	Transcribed — the core feature
Video messages	✅	Round video notes — transcribed
Photos	✅ for captions / --ocr	Scene-captioned and/or OCR'd; without it, photos show as a plain (photo)
Videos	optional, for --video-files	Regular videos — their audio is transcribed
Stickers	for --describe-hq	(sticker 😅) comes from JSON; tick to let --describe-hq caption the image too
GIFs	for --describe-hq	(animation) comes from JSON; tick to let --describe-hq caption it (multi-frame)
Files	⬜ not needed	Shown as (file: report.pdf) from the JSON metadata

⚠️ Drag the "Size limit" slider up. It defaults to 8 MB, and any file larger than that is not downloaded — those messages come out as [not exported]. Voice notes are tiny, but video notes, videos, and hi-res photos routinely exceed 8 MB, so raise the slider (toward the max) to be sure your media actually lands in the export. (This is the usual reason notes show as [not exported].)

You get a folder with a .json file plus voice_messages/, video_files/, photos/ … subfolders for whatever you ticked.

3. Run

From inside the export folder:

whispergram
# or, without installing:
python whispergram.py

…or point it at the folder:

whispergram "path/to/ChatExport_2026-06-20"

The result is merged_chat.md in the export folder.

Best quality (use your GPU)

Audio and video already use the most accurate model — Whisper large-v3 — on your GPU by default. For the best photo, sticker and GIF captions, just install the HQ extra — it's then used automatically, no flag — and, for speed, put torch on your GPU:

pip install -U "whispergram[describe-hq]"
# optional, for GPU-fast captions (match your CUDA, e.g. cu121/cu124):
pip install torch torchvision --index-url https://download.pytorch.org/whl/cu124
whispergram        # auto-uses large-v3 + Qwen2-VL; add --ocr --ocr-lang ukr+rus+eng for screenshot text

That runs large-v3 (audio/video) + Qwen2-VL (photos, stickers, GIFs). ⚠️ On Windows, a CUDA build of torch can clash with faster-whisper's GPU (cuDNN) — see GPU on Windows for the two reliable setups before installing CUDA torch.

Queue & resume

Pass several export folders to transcribe them back-to-back — the models load once and are reused, and sequential is safe for your GPU:

whispergram "ChatExport_Anastasia" "ChatExport_Olha" "ChatExport_Work" --out-dir "C:\merged"
# -> C:\merged\Anastasia.md, C:\merged\Olha.md, C:\merged\Work.md

Runs are resumable: each transcript/caption is cached to .whispergram_cache.json in the export folder as it's produced, so if you close the terminal or it crashes, just run it again — finished files are skipped and it continues where it left off. A progress bar shows done/total + ETA:

 60%|████████████        | 28/47 [02:14<01:31], audio_28.ogg

If two chats share a name, the second is saved as Work (2).md rather than overwriting the first, and a folder that fails (e.g. a corrupt export) is skipped so the rest of the queue still runs. --no-cache disables the cache; --out FILE sets a custom path for a single folder (it can't be combined with --out-dir).

---

Example output

[2026-06-20 12:33] Alex: did you get the files?
[2026-06-20 12:34] Alex (voice 6s): one sec, recording the summary now ...
[2026-06-20 12:35] You (photo, described): a screenshot of a calendar app | text: Sprint review - Fri 15:00
[2026-06-20 12:36] Alex (video-note 8s): [not exported]
[2026-06-20 12:36] You (sticker 😅)

Photo captioning is automatic once whispergram[describe] is installed; add --ocr for the in-image text, or --no-describe for a plain (photo) marker.

---

How each message appears

Message type	In the merged file
Text	[time] sender: message text
Voice note	[time] sender (voice 12s): <transcript>
Round video note	[time] sender (video-note 8s): <transcript>
Voice/video note with caption	[time] sender (voice 12s): <transcript> | caption: <text>
Voice/video not downloaded	[time] sender (voice 12s): [not exported]
Sticker	[time] sender (sticker 😅)
Photo, --no-describe	[time] sender (photo): caption (plain marker, no captioning)
Animation / GIF	[time] sender (animation)
Document	[time] sender (file: report.pdf): caption
Location / poll / contact	[time] sender (location) · (poll) · (contact)
Music / audio file	[time] sender (audio: Artist - Title) — transcribe with --audio-files
Regular video file	[time] sender (video) — transcribe the audio with --video-files
Photo (default, [describe] installed)	[time] sender (photo, described): a caption of the scene
Photo + --ocr	[time] sender (photo, described): <scene> | text: <text found in the image>
Photo + --ocr --no-describe	[time] sender (photo, text): <text found in the image>
Sticker / GIF + --describe-hq	[time] sender (sticker 😅, described): … · (animation, described): …

Markers can be turned off with --no-media-markers (voice/video notes are always transcribed).

---

Describe modes: photos, stickers & GIFs

Image captioning is opt-in via an extra. The best installed describer is used automatically — no flag needed:

Mode	How to enable	What it captions	Model	Size	Speed
Off	--no-describe	nothing (media shown as markers)	—	—	instant
Light	pip install whispergram[describe]	photos	BLIP-large	~1.9 GB	fast on CPU
High-quality (auto)	pip install whispergram[describe-hq]	photos + stickers + GIFs (GIFs multi-frame)	Qwen2-VL-2B	~4.4 GB	slow on CPU / fast on GPU

• Install the quality you want, then just run whispergram: if [describe-hq] is present it's used automatically (and captions stickers + GIFs); otherwise BLIP captions photos. --describe-hq forces HQ; --no-describe turns captioning off.
• HQ (Qwen2-VL) is markedly better on cartoons, characters and actions, and reads GIFs several frames at a time so it catches motion. BLIP is a quick photo gist (rough on cartoons).
• Add --ocr to also pull any in-image text. Everything is local; captions are best-effort, never literal fact. To run the models on your GPU, see GPU setup.

---

✅ Round-trip Validated

The merge has been validated against a real 770-message Telegram export (a live, audio-heavy chat — not a synthetic fixture). Every dimension was diffed against the source JSON:

Dimension	In export	In merged file	Result
Voice notes (downloaded)	4	4 transcribed	✅
Round video notes (not downloaded)	5	5 [not exported]	✅
Other media (stickers, photos, animations, videos, audio, …)	107	107 markers	✅
Text messages	654	654	✅
Messages dropped	—	0	✅

All 770 messages map to 770 lines — the per-type counts match the source exactly, and not-exported notes are never sent to the model. (An earlier version silently dropped 88 of those messages — every sticker, photo, and caption-less media item — leaving misleading gaps. The round-trip is what surfaced it.)

That export is private, so these counts were measured locally and are not reproducible from this repo. The synthetic export under tests/fixtures/ reproduces the same lossless mapping across every media type and guards it automatically in CI. A faithful merge is only proven once it has been run end-to-end and the output diffed back against every message type — structural validity alone is not enough.

---

Known Limitations

These follow from the Telegram export format and from speech recognition itself — not from a lack of effort in the tool:

Area	Status	Notes
Round video notes	Audio only, if downloaded	Telegram often excludes the binary; those show [not exported]
Music / audio_file	Off by default	Opt in with --audio-files; songs are otherwise not run through ASR
Photo OCR	Text-in-image only	--ocr reads visible text (great for screenshots), not a description of the scene; needs Tesseract + language packs
Photo descriptions	Best-effort, local	On by default with whispergram[describe] (BLIP via transformers) — captions are a short, English scene gist, not literal fact; --no-describe to skip
Stickers / GIFs / cartoons	--describe-hq only	Local models caption cartoons/memes roughly; --describe-hq (Qwen2-VL, multi-frame for GIFs) is much better but heavier — still best-effort, never exact
Speaker labels	Sender only	Each note is attributed to its Telegram sender; no in-audio diarization
Timestamps	Minute resolution	Telegram exports YYYY-MM-DDThh:mm; seconds are not shown
Reactions / edits / replies	Not represented	The merged file is a clean reading transcript, not a full forensic dump
Transcription accuracy	Model-dependent	large-v3 is best for uk/ru; --lang forces a language if auto-detect slips

---

Options

whispergram --device cpu --model large-v3-turbo   # no GPU, fast
whispergram --lang uk                             # force a language
whispergram --dry-run                             # preview the merge, no transcription
whispergram --audio-files                         # also transcribe music/long audio files
whispergram --video-files                         # also transcribe regular videos' audio
whispergram --ocr --ocr-lang ukr+rus+eng          # read text from photos (local Tesseract)
whispergram --no-describe                         # skip photo scene captions
whispergram --describe-hq                         # better captions + describe stickers/GIFs (Qwen2-VL)
whispergram --offline                             # zero network calls (use cached models only)
whispergram --out result.md                       # custom output path

Flag	Default	Notes
--device	cuda	cuda or cpu; auto-falls back to CPU if the GPU fails
--model	large-v3	try large-v3-turbo or medium if CPU is slow
--lang	auto	force a code like uk, ru, en if auto-detect mislabels
--batch-size	0	N>1 batches segments for a big GPU speedup; 0 = sequential (best quality)
--out	merged_chat.md	output file for a single folder (mutually exclusive with --out-dir)
--out-dir	off	collect each queued folder's transcript here as <chat name>.md
--no-cache	off	don't read/write the per-folder .whispergram_cache.json resume cache
--audio-files	off	also transcribe audio_file messages (music, long memos)
--video-files	off	also transcribe regular video files' audio track
--ocr	off	extract text from photos with local Tesseract OCR
--ocr-lang	eng	Tesseract language(s), e.g. ukr+rus+eng
--no-describe	off	skip photo scene captions (on by default when [describe] is installed)
--describe-model	blip-large	BLIP caption model id; use ...-base for faster/lighter
--describe-hq	off	high-quality describer (Qwen2-VL) + captions stickers/GIFs; needs [describe-hq]
--offline	off	use only cached models; make zero network calls
--no-media-markers	off	omit (sticker) / (photo) / (file) markers
--dry-run	off	map the chat without loading a model or transcribing
--setup-cuda-windows	—	copy CUDA DLLs next to ctranslate2, then exit (Windows GPU fix)

---

GPU (CUDA) setup

Linux / macOS: with a working CUDA install it runs as-is on --device cuda.

Windows — the common pitfall is RuntimeError: Library cublas64_12.dll is not found:

1. Install the CUDA runtime wheels (no full CUDA Toolkit needed):

   pip install nvidia-cublas-cu12 nvidia-cudnn-cu12
   pip install -U "ctranslate2>=4.5"
   

2. If it still can't find the DLL, copy them next to CTranslate2 (the reliable fix):

   python whispergram.py --setup-cuda-windows
   

3. Or skip the GPU entirely: --device cpu --model large-v3-turbo.

CTranslate2 loads cuBLAS/cuDNN lazily in native code that ignores os.add_dll_directory, which is why placing the DLLs inside the package dir is the dependable solution.

GPU for photo/sticker/GIF captions is separate from Whisper's. The describe models (BLIP / Qwen2-VL) use PyTorch, and pip install fetches the CPU build by default — so captioning runs on the CPU even when Whisper is on your GPU. For fast captioning (especially --describe-hq), install a CUDA build of torch:

pip install torch torchvision --index-url https://download.pytorch.org/whl/cu121   # match your CUDA

whispergram auto-detects CUDA and moves the caption model to the GPU — no flag needed.

⚠️ Windows: a CUDA torch can clash with Whisper-on-GPU. A CUDA build of torch bundles its own cuDNN, which can collide with the cuDNN that faster-whisper (CTranslate2) uses — surfacing as OSError: [WinError 127] … cudnn_*.dll on startup. Both can't reliably share the GPU out of the box, so pick one of these stable setups:

• Whisper on GPU + captions on CPU (default, recommended): keep the CPU build of torch — pip install torch torchvision --index-url https://download.pytorch.org/whl/cpu. Fast audio, slower captions.
• Captions on GPU + Whisper on CPU: pip uninstall nvidia-cudnn-cu12 nvidia-cublas-cu12, install a CUDA torch, and run with --device cpu. Fast captions, slower audio.

whispergram prints this guidance if it hits the conflict.

Both on the GPU — two passes (fastest for big batches). Because runs are resumable, you can do each heavy step on the GPU in turn without the two libraries ever colliding:

# Pass 1 — transcribe everything on the GPU (CPU torch + faster-whisper GPU), no captions:
whispergram <folders> --out-dir DIR --no-describe

# switch torch to CUDA (captions-on-GPU setup):
pip uninstall -y nvidia-cudnn-cu12 nvidia-cublas-cu12
pip install torch torchvision --index-url https://download.pytorch.org/whl/cu124   # match your CUDA

# Pass 2 — caption on the GPU; Whisper runs on CPU but every transcript is already cached, so it's
# instant and only the captioning does work:
whispergram <folders> --out-dir DIR --describe-hq --device cpu

Both expensive stages run on the GPU, the resume cache means no transcription is repeated, and the cuDNN clash never happens because only one library touches the GPU per pass.

Ukrainian / Russian OCR

--ocr needs the Tesseract binary (auto-found on Windows since 0.8.2) and the language packs. On Windows: winget install UB-Mannheim.TesseractOCR, then add the ukr/rus data — either re-run that installer and tick them, or drop ukr.traineddata + rus.traineddata (tessdata_best) into a folder and point TESSDATA_PREFIX at it. Verify with tesseract --list-langs.

---

FAQ

How do I transcribe Telegram voice messages? Export the chat from Telegram Desktop as JSON (with voice messages), then run whispergram in the export folder. Every voice note is transcribed with Whisper and merged into the text chat.

Is it private / offline? Does my audio leave my machine? Yes. Transcription, captioning and OCR run locally and need no account or API key. The tool makes no network calls with your data — your chat audio, photos and transcripts never leave your machine. The only network use is a one-time download of the model weights (public files) from Hugging Face; usage telemetry is off by default, and --offline forces cache-only with zero network calls once the models are downloaded.

Do I need a GPU? No. It runs on CPU (--device cpu); use --model large-v3-turbo for speed. A CUDA GPU is faster.

Does it handle round video messages / video notes? Yes — round video_message notes are transcribed from their audio, just like voice notes. Regular video files are transcribed too with --video-files.

Can it read text from photos / screenshots? Yes — --ocr runs local Tesseract over photos and drops the extracted text inline as (photo, text): ... (ideal for screenshots).

Can it describe what's in a photo, not just the text? Yes, and it's automatic: once you pip install whispergram[describe], photos are captioned by a small local model (BLIP via transformers — uses your GPU if you have one, else CPU) with no flag needed. It composes with --ocr to give both the scene and the in-image text. Captions are a short, English, best-effort gist. Pass --no-describe to turn it off, or --describe-model Salesforce/blip-image-captioning-base for a faster/lighter model. The BLIP-large model (~1.9 GB) downloads once on the first photo, then stays offline.

Can it describe stickers and GIFs too? Yes, with --describe-hq (pip install whispergram[describe-hq]). That switches to a stronger model (Qwen2-VL) that's much better on cartoons and actions, and it reads GIFs multi-frame to catch the motion — e.g. (animation, described): a character in a suit walking into an arena. It's heavier (~4.4 GB; slow on CPU, fast on GPU), and cartoon/meme captions are still best-effort, never exact.

Which languages work? Any language Whisper supports. large-v3 handles Ukrainian and Russian well; use --lang uk (or ru, en, …) to force one if auto-detection slips.

How is this different from Telegram Premium's transcription? Premium transcribes one message at a time, by hand, in the app. whispergram transcribes the entire chat in one pass, offline, and produces a single searchable file.

---

Project Structure

whispergram/
├── whispergram.py             # The tool: text reconstruction, mapping, transcription, CLI
├── requirements.txt           # Runtime dependency (faster-whisper)
├── pyproject.toml             # Packaging + ruff + pytest configuration
├── CHANGELOG.md
├── CONTRIBUTING.md
├── LICENSE
├── README.md
│
├── .github/
│   ├── workflows/
│   │   ├── ci.yml             # ruff + pytest on Python 3.9–3.13 (no transcription deps)
│   │   └── publish.yml        # tag v* → verify version → build → PyPI (trusted publishing)
│   ├── ISSUE_TEMPLATE/
│   └── dependabot.yml
│
└── tests/
    ├── test_whispergram.py    # 85 offline tests — no model download or GPU required
    └── fixtures/
        └── sample_export/
            └── result.json    # synthetic export (safe to commit; used by tests + CI)

---

⚠️ Privacy

This tool processes private conversations, and the transcripts it produces are just as sensitive as the audio. Two rules:

• Nothing leaves your machine. Transcription, captioning and OCR are fully local; the tool makes no network calls with your data and needs no credentials. The only network use is a one-time download of public model weights from Hugging Face — telemetry is off by default, and --offline guarantees zero network calls once the models are cached.
• Never commit your exports or transcripts. The included .gitignore blocks chat data (*.json, audio files, merged_chat.md, ChatExport_*/) by default — keep it. Build your repo in a folder separate from any export, keep any --out path inside the export folder, and run git status before pushing to confirm only code is staged. The only data file in this repo is the synthetic fixture under tests/fixtures/.

---

Requirements

• Python 3.9+
• ffmpeg on your PATH
• faster-whisper >= 1.0 (pip install -r requirements.txt)
• For NVIDIA GPU on Windows: nvidia-cublas-cu12, nvidia-cudnn-cu12, ctranslate2>=4.5
• For --ocr (optional): the Tesseract binary on your PATH (with language packs, e.g. ukr, rus) plus pip install whispergram[ocr]
• For photo descriptions (optional): pip install whispergram[describe] (transformers + torch — prebuilt wheels, no compiler; uses your GPU if present). Captioning is then automatic; the ~1.9 GB BLIP-large model downloads once on the first photo, then runs offline. Use --describe-model Salesforce/blip-image-captioning-base for a lighter model, or --no-describe to turn it off
• For high-quality captions + sticker/GIF describe (optional): pip install whispergram[describe-hq] (adds torchvision) and pass --describe-hq. Uses Qwen2-VL (~4.4 GB, slow on CPU / fast on GPU)

The test suite needs none of the above — only ruff and pytest.

---

Changelog

See CHANGELOG.md for the full version history.

Contributing

See CONTRIBUTING.md for the development setup, the privacy rule, and the versioning / release policy.

License

MIT