Repository navigation
Releases
Releases · phonology024/babelscribe
v0.4.0
New
babelscribe doctor: checks your setup and reports what is missing.--prompt "names, terms": custom vocabulary (vocabularyin the MCPtranscribetool).--translate: translates speech to English (translatein MCP).--max-line-chars: subtitle line wrapping, with a smaller default for Thai and CJK.-f ass: karaoke subtitles that highlight each word as it is spoken.- New
vietnamese-toantamoption for--text-model. It has no FLEURS benchmark yet, so--accuratedoes not use it.
Fixed
- Names with dots keep them, so
meeting.v1.mp4andmeeting.v2.mp4no longer overwrite each other's subtitles. -ocreates its output folder before the slow work starts.- An unknown or non-numeric
--devicenow fails straight away and lists the devices that exist.
Behaviour changes
-ois a base path and the extension is now always appended:-o x.srtwritesx.srt.srt. Use-o x.
Other
- CI now covers Python 3.9-3.13 on Windows, Linux and macOS.
- Added
API_STABILITY.md, and a Scoop manifest inpackaging/scoop(not yet submitted to a bucket).
Install: pip install -U babelscribe
v0.3.2
CI: publish to the MCP Registry via GitHub OIDC on each tag Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
v0.2.1
Portable x64 binaries; self-heal a whisper-cli that crashes on this CPU CI built whisper-cli with GGML_NATIVE, so it used the runner's AVX-512 and died with STATUS_ILLEGAL_INSTRUCTION on most user PCs (seen on an Intel 12th gen + RX 9070 XT: 'devices' fell back to CPU). x64 builds now target an AVX2 baseline. A cached binary that crashes this way is deleted and downloaded again once, with a clear error if it still cannot run. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
v0.2.0
Add --accurate mode and 16-language FLEURS benchmark - --accurate: large-v3 + beam 5, or the best per-language fine-tune in hybrid mode (Thai: Pathumma Whisper, Hindi: vasista22 large-v2), chosen from bench/fleurs.py results - bench/fleurs.py: FLEURS benchmark with Whisper-style normaliser, number spelling, edge-trimmed score, saved hypotheses - bench/bench.py: --accurate mode on the TED talks - fix bf16 conversion patch string; test that every module imports - README: FLEURS table, updated TED table; version 0.2.0 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
v0.1.1
CI: build Linux CUDA in the NVIDIA container, drop Windows CUDA (Vulk…
Pre-converted fine-tunes (models-1)
Pre-converted community Whisper fine-tunes in whisper.cpp ggml format (q8_0), so --accurate works without installing PyTorch.
babelscribe downloads these automatically and checks their SHA-256 (SHA256SUMS.txt). This is not a software release.
| File | Original model | Licence | Authors |
|---|---|---|---|
| ggml-thai-pathumma-q8_0.bin | nectec/Pathumma-whisper-th-large-v3 | Apache-2.0 | NECTEC |
| ggml-thai-thonburian-q8_0.bin | biodatlab/whisper-th-large-v3-combined | Apache-2.0 | biodatlab (Thonburian Whisper) |
| ggml-hindi-vasista-q8_0.bin | vasista22/whisper-hindi-large-v2 | Apache-2.0 | vasista22 (Speech Lab, IIT Madras) |
Modifications (Apache-2.0 §4b): converted from Hugging Face Transformers weights to ggml with whisper.cpp's
convert-h5-to-ggml.py (bf16 weights loaded as float32) and quantized to q8_0 with whisper-quantize. No retraining.
Each file remains under its original Apache-2.0 licence (https://www.apache.org/licenses/LICENSE-2.0); all credit for the
models goes to their authors. The tokenizer files come from openai/whisper (MIT).