Verbatim speech-to-text with word timestamps on a 4 GB GPU. CrisperWhisper 2.0 Turbo (CTranslate2, CUDA int8) writes down exactly what was said (fillers, false starts, repetitions kept); a wav2vec2 CTC forced alignment then gives every word its start and end time. Output: text, word JSON, SRT and VTT. Built and measured for Turkish, works for every language the models know (docs/diller.md). Linux or Windows via WSL2; one command to install.

Türkçe README · Languages and word splitting · Measurements · Install guide for AI agents · Responsible use

Model license: the CrisperWhisper 2.0 weights are non-commercial (nyra health Non-Commercial Research License). This repository's code is Apache-2.0, but running it inside a paid product needs a commercial license from Nyra Health. Weights are not in this repo; the installer downloads them from a pinned commit. See NOTICE.

cikti/kayit_20261023_0210/
  transkript.txt      verbatim text
  kelimeler.json      every word: start, end, score, how it was timed; language; timings; both models' identities
  transkript.srt/vtt  caption lines (sentence or ~7 s)
  kelimeler.srt/vtt   one cue per word (karaoke-style)
  ses_16k_mono.wav    the audio as the models heard it

kelimeler.json word entry: {"kelime": "dünya.", "start": 0.62, "end": 1.01, "skor": 0.93, "yontem": "ctc"}. yontem is ctc for a real alignment, ara_deger (interpolated between neighbours) when the alignment model's vocabulary had none of the word's characters, esit_bolme when a segment could not be aligned at all.

Needs: NVIDIA GPU with ≥ 4 GB VRAM and driver ≥ 570 (CUDA 12.8), Python 3.10-3.12, curl, ~7 GB disk (torch 3 GB, model 1.6 GB download + 1.6 GB converted, alignment models 0.4-1.3 GB per language). No sudo, no system CUDA install, no PATH changes; everything goes under the root you choose.

Linux

git clone https://github.com/alexturkey/crisperwhisper-turkce-kelime-zamani.git
cd crisperwhisper-turkce-kelime-zamani
bash scripts/kur.sh --kok ~/stt            # ~15 min, ~6 GB download; markers PYTHON_OK MODEL_OK DONUSUM_OK HIZALAMA_OK AYAR_OK KONTROL_OK KURULUM_OK
.venv/bin/python -m transkript ornek/ornek.mp3 --dil tr

Windows (the CTranslate2 fork ships Linux wheels only, so the engine runs inside WSL2; the wrappers hide that)

wsl --install -d Ubuntu          # once, admin; reboot; then open "Ubuntu" once to create your user
git clone https://github.com/alexturkey/crisperwhisper-turkce-kelime-zamani.git
cd crisperwhisper-turkce-kelime-zamani
powershell -NoProfile -ExecutionPolicy Bypass -File scripts\kur.ps1 -Kok E:\stt
powershell -NoProfile -ExecutionPolicy Bypass -File scripts\stt.ps1 -Ses ornek\ornek.mp3 -Dil tr

-ExecutionPolicy Bypass applies to that one process only; read scripts\kur.ps1 and scripts/kur.sh before running them, as with any installer. Re-run after a failure: verified downloads are skipped.

--diller "tr en ja" (kur.sh) / -Diller "tr en ja" (kur.ps1) chooses which alignment models to pre-download (default tr en); any other language in the table downloads on first use.

.venv/bin/python -m transkript kayit.mp3                       # Turkish (default), output under cikti/
.venv/bin/python -m transkript kayit.wav --dil en --cikti out  # English
.venv/bin/python -m transkript a.mp3 b.wav c.mp3 --cikti out   # batch: models load once, out/<name>/ per file
.venv/bin/python -m transkript --liste isler.tsv --cikti out   # batch from a list: <path>[<TAB><lang>] per line
.venv/bin/python -m transkript kayit.m4a --dil auto            # detect on the first 30 s (needs ffmpeg for m4a/mp4)
.venv/bin/python -m transkript --diller                        # languages with an alignment model (tested / experimental)
.venv/bin/python -m transkript --kontrol                       # environment check -> KONTROL_OK

Input: anything libsndfile decodes (wav, flac, ogg, mp3); video or other containers need ffmpeg on PATH. Options: --bicim json,srt,vtt, --hizalama-modeli <HF wav2vec2 CTC id> (other language or your own checkpoint), --hizalama-cihazi cpu (alignment on CPU if the GPU is too small; slower), --zaman-asimi.

1. Audio → 16 kHz mono WAV (soundfile + soxr; ffmpeg only for video containers).
2. ASR worker (own process): crisperwhisper CT2 backend, compute_type=int8, verbatim mode, longform "continuation" chunking with hallucination repair and early-EOT recovery, native word timestamps on. The process exits and frees the GPU.
3. Alignment worker (own process): the ASR words are grouped into ≤ 18 s segments (sentence ends, > 2.5 s gaps); each segment's characters are force-aligned with the language's wav2vec2 CTC model (torchaudio.functional.forced_align + merge_tokens), so a word's time spans from its first to its last recognised character; silence is never swallowed into a word. CJK/Thai text is timed per character (docs/diller.md).
4. Checks (chronological, inside the audio, no overlaps) and outputs. The native CrisperWhisper times stay in kelimeler.json → asr_kelimeleri for comparison.

Two processes instead of one so a 4 GB card never holds both models: CT2 turbo int8 ≈ 1371 MB peak, the Turkish alignment model ≤ 1653 MB peak (nvidia-smi, GTX 1050 Ti).

--dil xx for any Whisper language (--dil auto detects from the first 30 s). What a "word" is depends on the script (full table and reasoning: docs/diller.md):

No bundled morpheme splitter: if you need Japanese words (MeCab/fugashi), Chinese words (jieba) or Thai words (pythainlp), run the tokenizer on metin and regroup the per-character times (first char start → last char end); docs/diller.md shows the 10-line recipe. Every word carries yontem: ctc (aligned), ara_deger (interpolated between neighbours when a character is missing from the model's vocabulary), esit_bolme (the whole segment could not be aligned; its words share the segment time equally).

GTX 1050 Ti 4 GB, WSL2 Ubuntu 24.04, i7-7700HQ. Method and the full table: docs/olcumler.md.

Batch mode, all 12 items in one run: 602 s wall clock, of which about 480 s is importing torch and loading models from a Windows drive mounted into WSL2. Single-file runs pay that loading every time (459-727 s per file here), so pass several files or --liste. Peak VRAM: 1371 MB (ASR) / up to 2411 MB (alignment).

How "normalised" is computed (tests/wer_olc.py, applied the same way to reference and our output; both numbers are always printed):

1. lowercase (Turkish I/İ aware), punctuation and [event] tags removed (this alone is the raw number);
2. circumflex folded: â/î/û → a/i/u (hâlâ = hala);
3. numbers written out in Turkish words, suffix kept (40.000'in = kırk binin);
4. apostrophe removed, suffix joined (Okulu'ndan = okulundan);
5. ile joined to the previous word (yöntemi ile = yöntemiyle).

Nothing else is forgiven: names, wrong words and compound spacing (çevrimiçi → çevrim içi) still count.

The synthetic rows are clean TTS voices: they show the pipeline works end to end, not how it does on real or noisy recordings. FLEURS (Google, tr_tr test split, first 10 clips, CC BY 4.0) is real read speech; ten clips is a smoke test, not a benchmark. Word-time accuracy was not measured against hand labels; the method's behaviour is described in docs/diller.md. Nyra reports CrisperWhisper 2.0's numbers on their own benchmark (github.com/nyrahealth/CrisperWhisper); those are theirs, not ours.

• Non-commercial model license (above). Verbatim output is personal data: see ETHICS.md.
• One GPU job at a time; ~4.5 GB free RAM during conversion and alignment.
• Windows needs WSL2 (no native Windows path for the CT2 fork). Linux + CUDA 12 only; no AMD/Intel, no CPU path for the ASR (CTranslate2 int8 on CPU runs but we did not measure it; set "cihaz": "cpu" in ayarlar.json to try).
• Long silence or music can still produce repeated or invented text; the library's repair catches most of it, not all.
• Overlapping speakers, no diarization. Languages other than tr/en/ja are experimental (docs/diller.md).

Code: Apache-2.0, Copyright 2026 Ali Haydar Ozturk. Models: CrisperWhisper 2.0 Turbo (Nyra Health, non-commercial research license; inference code MIT), wav2vec2 alignment checkpoints (mpoyraz CC BY 4.0, jonatasgrosman Apache-2.0, torchaudio/fairseq MIT), CTranslate2 (OpenNMT, MIT), PyTorch. Full list: THIRD_PARTY.md. Sister project: voxcpm2-turkce-ses-tasarimi (Turkish voice design, the source of the example audio). Not affiliated with Nyra Health or OpenNMT.