WhisperDRZ is a speaker-aware automatic speech recognition model. It transcribes audio into text with word-level timestamps, per-line speaker tags, and non-speech event tags. It is a Whisper-style encoder-decoder model and handles long audio by chunking and stitching internally.
This repository is inference only. Weights are distributed on the Hugging Face
Hub at fluxions/whisperdrz.
🎙️ Try it in the browser — no install needed: fluxions.ai/transcribe.
📖 Full write-up — how it works, the timing-token and diarization approach, honest results (WER, DER, timing, events, multilingual), and what didn't work: writeup.md.
pip install -e . # add ".[demo]" for the gradio demo, ".[eval]" for the metricsRequires Python 3.12+ and FFmpeg (for audio decoding: apt install ffmpeg /
brew install ffmpeg). Runs on a CUDA GPU or on CPU.
Flash Attention is optional: when flash-attn is installed (and a CUDA GPU is
available) it is used with CUDA graphs for fast decoding; otherwise a pure-torch
attention path is used automatically. flash-attn needs torch present at build
time, so install it separately: pip install flash-attn --no-build-isolation.
whisperdrz audio.wav # defaults: --model whisperdrz-large-v3.safetensors, --lang en
whisperdrz audio.wav --output_format json > out.json
whisperdrz audio.wav --model my-checkpoint.pt --lang auto # override the defaults--model defaults to whisperdrz-large-v3.safetensors and accepts a local checkpoint,
a filename hosted in the weights repo, or a Hugging Face repo id. The weights
download automatically on first use. --lang defaults to en; use auto to detect.
import whisperdrz
from whisperdrz.audio import load_audio, SAMPLE_RATE
transcriber = whisperdrz.load_model("whisperdrz-large-v3.safetensors", lang="en")
audio, _ = load_audio("audio.wav", sample_rate=SAMPLE_RATE)
result = transcriber.transcribe(audio.mean(0)) # mono, 16 kHz
print(result.text) # speaker-tagged text with timestamps
print(result.segments) # list of {speaker, start, end, text}Each line begins with a speaker tag. Timed words and tags are wrapped in a start/end timestamp pair; not every word is timed, but the first and last word of each line always are:
[0] <|0.00|>Hello<|0.45|> there <|0.80|>world.<|1.10|>
[1] <|1.20|>Hi<|1.40|> <|1.45|>[laugh]<|1.60|> <|1.70|>there.<|1.95|>
[0],[1], ... are speaker IDs;[c]marks crowd/ambient.<|t|>are timestamps in seconds (two decimals), always in a pair wrapping a word or tag.- The first and last word of every line are always timed; middle words may be bare.
[laugh],[breath], and similar are non-speech event tags.
transcribe() returns a TranscribeResults with the raw text plus parsed
segments, each a dict of speaker, start, end, and text.
WhisperDRZ reports transcription, diarization, and joint metrics:
- WER (word error rate) — Levenshtein word distance after normalization (whisper-normalizer), measuring transcription accuracy, speaker-agnostic.
- WDER (word diarization error rate) — of the words that align between reference and hypothesis, the fraction assigned to the wrong speaker under the best speaker permutation. It isolates diarization quality from WER (insertions/deletions don't count toward WDER).
- cpWER (concatenated minimum-permutation WER) — concatenate each speaker's words into one stream, pick the speaker permutation that minimises total word errors, then report a single WER over the concatenated streams. Unlike WDER it does charge insertions/deletions, so it jointly scores transcription and diarization in one number (computed with meeteval).
- tcpWER (time-constrained cpWER) — cpWER with the extra constraint that a hypothesis word only matches a reference word if their timestamps fall within a collar, so it also rewards accurate word timing (meeteval).
- DER (diarization error rate) — time-based: missed speech + false alarm + speaker confusion, over total reference speech, at a 0.25s collar. The standard diarization metric; speaker-count- and word-agnostic.
Measured on the released checkpoint:
| Benchmark | WER | DER (miss / FA / conf) | cpWER / tcpWER |
|---|---|---|---|
| ESB (English ASR, 1000 utts) | 9.6% macro / 5.9% micro | — | — |
| Internal conversational (26 clips) | 11.1% | — (WDER 33%) | 46% cpWER |
| VoxConverse dev (216, overlap-heavy) | — | 26.3% (3.1 / 15.8 / 7.5) | — |
| CALLHOME eng (140, 2-spk telephone) | — | 38.6% (5.9 / 14.0 / 18.7) | — |
| AMI test (16 meetings, Mix-Headset) | 22.8% | 50.6% (10.6 / 28.9 / 11.1) | 72% / 84% |
DER is at the standard 0.25s collar, scoring overlapping speech. AMI cpWER/tcpWER and the 22.8% WER use oracle speaker count (the model is told how many speakers each meeting has); its natural clustering over-fragments long meetings (up to ~35 speakers where there are 4), which makes the permutation-based cp-metrics intractable. Interestingly, oracle-count raises AMI DER to 64.9% (confusion 26.6%): the natural over-fragmentation was quietly flattering DER, whose optimal mapping absorbs excess clusters as false alarm while keeping confusion low.
WhisperDRZ is an ASR-first model. Transcription is strong across the board (AMI Mix-Headset WER 22.8%), but diarization trails purpose-built systems (~10–25% DER) — confusion dominates, and the ~49-point gap between AMI WER (23%) and cpWER (72%) is almost entirely speaker attribution: the words are right, but who said what is often wrong, especially on long, overlapping multi-party audio. See the write-up for full analysis (timing, non-speech events, multilingual).
A note on false alarm — WhisperDRZ has no VAD. It transcribes over silence and music, so false alarm is a large part of the DER above (28.9% of the 50.6% on AMI). Gating the hypothesis to detected speech before scoring — a component the shipped model doesn't include — recovers much of it. The DER above is the honest out-of-the-box number; the columns below show what a non-speech filter would buy:
| Benchmark | DER (ungated) | + silero VAD | + oracle VAD |
|---|---|---|---|
| VoxConverse dev | 26.3% | 24.0% | 20.9% |
| CALLHOME eng | 38.6% | 34.2% | 28.6% |
| AMI test | 50.6% | 40.6% | 31.1% |
silero VAD is a real off-the-shelf detector; oracle VAD (reference speech regions) is the ceiling. What a VAD cannot fix is speaker confusion and over-hypothesized overlap — the genuine diarization error that remains after detection is perfect.
Note: an earlier revision reported VoxConverse dev DER as 40.7%. The 26.3% above is a re-measurement on the released checkpoint with the current scoring pipeline (pyannote.metrics, 0.25s collar, overlap scored); the discrepancy is still being reconciled.
Reproduce WER/WDER/cpWER on your own data with the metrics in whisperdrz.evals
and the runner:
pip install -e ".[eval]"
# manifest.jsonl: one {"audio": "a.wav", "text": "[0] ref ... [1] ..."} per line
python scripts/eval.py manifest.jsonl --model whisperdrz-large-v3.safetensorstext is the reference transcript; [N] speaker tags are optional and WDER /
cpWER are only computed for multi-speaker references. DER and tcpWER require
time-aligned references (RTTM / word timings) and the meeteval + pyannote.metrics
tooling used for the benchmarks above.
python demo/app.py --model whisperdrz-large-v3.safetensors@software{whisperdrz_2026,
author = {Coultas Blum, Harry},
month = {07},
title = {{WhisperDRZ}},
url = {https://github.com/fluxions-ai/whisperdrz},
version = {0.1.0},
year = {2026}
}
MIT. See LICENSE.