- Category
- AI Tools
- Rank
- No. 2024Tools index
- Pricing
- Open Source
- Type
- TOOL
- GitHub
- 4 stars
- Latest release
- v0.1.0
- Date
About
A transcription tool built on the Rumble-2-Flash speech model that produces word-level timestamps, optional speaker labels and subtitles. The model's decoder writes the text, and its CTC head force-aligns that text to the audio using the encoder pass already computed, so alignment does not need a second model. Output formats include JSON, SRT, VTT, TSV and TXT.
At a glance
rumblex is a Python command-line tool and library that turns audio or video files into transcripts with per-word timestamps, optional speaker labels and subtitle files (SRT, VTT, JSON, TSV, TXT). It runs the open Reson8/Rumble-2-Flash speech model locally and reuses that model's built-in CTC head to time each word, so alignment needs no second model. Worth trying if you subtitle or index recordings on your own GPU and want a WhisperX alternative.
Good fit
Developers and researchers batch-transcribing meetings, calls or interviews who need word-level timing or speaker turns. In the author's 0.1.0 runs on one RTX 4090, it was 3 to 5 times faster than WhisperX large-v3 and had lower word error on AMI meetings (10.9% vs 17.2%) and Earnings-22 calls (9.3% vs 10.6%).
Consider something else when
Your audio is mostly clean read English or Dutch and accuracy matters more than speed: in the same runs WhisperX had lower error on LibriSpeech (1.6% vs 2.2%) and FLEURS Dutch (7.0% vs 9.2%). Also skip it for Chinese or Japanese, where space-based word splitting gives one "word" per chunk, or if you have no GPU.
Start here
With Python 3.11 or 3.12 and ffmpeg on your PATH, run pip install rumblex, then rumblex talk.wav -f srt vtt --max_line_width 42. Success is SRT and VTT files in the current folder with cues split at 42 characters. Next, add --language, a --prompt with names or jargon as hotwords, or --diarize after installing rumblex[diarize].
Cost and requirements
Free under Apache-2.0, as is the Rumble-2-Flash model; pyannote's models have their own terms. Needs Python 3.11 or 3.12, PyTorch 2.6 or later and ffmpeg. A GPU is strongly recommended: the model runs in float32 on a 6 GB GTX 1060 at the default batch size of 8, and CPU works but is slow. pip's default PyTorch is CPU-only on Windows, so install a CUDA build first there. RTX 50-series cards need CUDA 12.8 or later.
Limits to know
Version 0.1.0 is marked Alpha. Word timing is limited to 13 encoder frames per second, so word ends are less precise than WhisperX (57 vs 34 ms mean error on 200 LibriSpeech utterances, measured against another forced aligner). Each chunk of up to 20 seconds is decoded alone, so sentences can break at chunk boundaries. Speaker labels require accepting gated pyannote model terms and a Hugging Face login. The faster vLLM backend needs Linux and a supported GPU.
Still unclear
- How accurate is it on any of the 45 languages the model card lists other than English and Dutch? Only those two were benchmarked.
- How much does the WER lead on meetings and calls hold up on long recordings, where both tools segment audio with a VAD?
Sources checked 2026-11-21. Source-based assessment.
Tags
Tech Stack
Comments (0)
No comments yet
Editorially curated, with community endorsements as a secondary signal. Corrections welcome.