v0.2

Full Changelog: v0.1.2...v0.2

• improved progress bar with more information
• improved file selector

v0.1.2

Full Changelog: v0.1.1...v0.1.2

Windows executable now works again.

v0.1.1

Full Changelog: v0.1.0...v0.1.1

Esperanto-STT v0.1.1

Esperanto speech-to-text: turn mp3 audio into .srt subtitle files, with a small desktop GUI. Runs on Linux and Windows, CPU-only, fully offline.

Special thanks to all the contributors of the Esperanto Dataset of the Common Voice Project that made this possible.

Features

Transcription

• INT8 ONNX export of NVIDIA's Esperanto Conformer-Transducer Large (~120M params) via onnx-asr + ONNX Runtime — no PyTorch, NeMo or CUDA
• Word-level timestamps → properly timed subtitle blocks
• Input formats: mp3, wav, ogg, flac (m4a/aac only when ffmpeg is installed)
• Long files are transcribed in 30 s chunks (batched) at ~10-15× realtime on CPU — a 48-minute podcast takes ~3-5 minutes
• Output is the model's raw text: lowercase Esperanto with diacritics (ĉĝĥĵŝŭ), no punctuation

Subtitles

• The .srt is written incrementally — created when transcription starts and growing block by block, so you can already read it (editor, player, preview window) while the progress bar runs; a partial file is kept if a run fails midway
• Readable blocks: ≤ 84 chars (2×42) and ≤ 7 s per block, split on pauses > 1.5 s, min 0.5 s per block

GUI (in Esperanto)

• Dark-mode desktop app (CustomTkinter), fully in Esperanto with diacritics
• File list with per-file progress; an srt button per file opens a read-only preview — works mid-run, too
• When a run completes, a preview window opens automatically for the last file, with a kopii ĉion (copy all) button
• OS-native file picker (zenity on GNOME/Linux, native dialog on Windows)

CLI

• eo-stt episode.mp3 more.mp3 -o subtitles/ — one .srt per input, written next to the file or into -o
• Chunk-level progress; ASCII-only output (safe on any Windows console)

Portable, offline builds

• Self-contained executables for Linux and Windows — no Python, no admin rights, no internet needed on the target machine
• The INT8 model (~142 MB) is bundled beside the executable (model/) — the app never downloads anything at runtime
• One executable, two interfaces: no arguments → GUI, with arguments → CLI
• Point the app at a different model directory with the EO_STT_MODEL_DIR environment variable (or --model on the CLI)

Downloads

Artifact	Platform
eo-stt-v0.1.1-linux.tar.gz	Linux (x86_64)
eo-stt-v0.1.1-windows.zip	Windows (x64) - broken in this release

Both artifacts are built automatically by CI (GitHub Actions) from the v0.1.0 tag; the model is fetched at build time (pinned commit, sha256-verified) and bundled.

Unpack, then from the eo-stt/ folder run ./eo-stt for the GUI or ./eo-stt episode.mp3 -o out/ for the CLI. On Windows: unzip and double-click eo-stt.exe for the GUI, or run it from a terminal for the CLI.

Notes & known limitations

• Model output is lowercase without punctuation — passed through as-is (that's the model, not a bug)
• Audio is chunked in hard 30 s blocks without overlap: a word straddling a chunk boundary can rarely come out garbled or duplicated
• Windows Defender may flag PyInstaller-built executables as a false positive — add an exclusion for the app folder if that happens
• Build from source: see the README (python fetch_model.py, python build.py; tests: python -m pytest)

License

• Code: CC0 1.0 Universal (public domain dedication)
• Bundled model: CC-BY-4.0 — NVIDIA's Esperanto Conformer-Transducer Large via OpenVoiceOS/nvidia-eo-conformer-transducer-large-onnx (attribution kept in the README)

Big parts of this repo have been written by Mistral Larg 4 using OpenCode.

v0.1.0

Full Changelog: https://github.com/BitsAndDopamine/Esperanto-STT/commits/v0.1.0

Esperanto-STT v0.1. -- broken release

This release can't open files. Use v0.1.1 instead

License

• Code: CC0 1.0 Universal (public domain dedication)
• Bundled model: CC-BY-4.0 — NVIDIA's Esperanto Conformer-Transducer Large via OpenVoiceOS/nvidia-eo-conformer-transducer-large-onnx (attribution kept in the README)