faster-whisper
faster-whisper: 4× faster than reference Whisper using CTranslate2 — production sweet spot. Skip the install: WhipScribe transcribes it in your browser. Packs: $4 for 500 minutes, $12 for 2,000, $24 for 5,000; Workspace $19 a month. Up to 10 hours or 5 GB per file. Measured turnaround: about 2 minutes per audio hour. 99 languages. Checked 25 Oct 2026.
Don’t want to install faster-whisper? Drop the file here.
Every recording starts with a free preview of its opening minutes. Credits never expire.
Production-grade Whisper inference. CTranslate2 rebuild, ~4× faster than reference openai-whisper at the same WER, smaller VRAM, drop-in Python API. Pick a model, copy a recipe.
Install faster-whisper and run it on CPU or an NVIDIA GPU
faster-whisper 1.2.1 is the current PyPI release (31 Oct 2025). It needs Python 3.9 or newer and, unlike openai-whisper, no system FFmpeg: PyAV bundles the FFmpeg libraries.
The README names CUDA 12 only (checked 26 Oct 2026). Word timestamps: pass word_timestamps=True to transcribe(). VAD: vad_filter=True runs Silero VAD, which by default removes only silences longer than 2 seconds, and batched transcription turns it on by default.
Install from PyPI.
pip install faster-whisperTranscribe. Use float16 on a GPU, int8 on a CPU.
from faster_whisper import WhisperModel model = WhisperModel("large-v3", device="cuda", compute_type="float16") # CPU: WhisperModel("small", device="cpu", compute_type="int8") segments, info = model.transcribe("audio.mp3", beam_size=5, vad_filter=True) for s in segments: print("[%.2fs -> %.2fs] %s" % (s.start, s.end, s.text))GPU: the latest ctranslate2 needs cuBLAS and cuDNN 9 for CUDA 12. On Linux both can come from pip; set LD_LIBRARY_PATH as the README shows before starting Python.
pip install nvidia-cublas-cu12 nvidia-cudnn-cu12==9.*Stuck on CUDA 12 with cuDNN 8? The README's fix is an older ctranslate2 (3.24.0 for CUDA 11 with cuDNN 8).
pip install --force-reinstall ctranslate2==4.4.0
Sources:
- PyPI · faster-whisper (1.2.1, 31 Oct 2025), fetched 26 Oct 2026
- SYSTRAN/faster-whisper · README (GitHub), fetched 26 Oct 2026
faster-whisper model names, and where large-v3-turbo comes from
WhisperModel("<name>") downloads a converted CTranslate2 model from Hugging Face. The name map in faster_whisper/utils.py, read on 26 Oct 2026:
There is no Systran/faster-whisper-large-v3-turbo: Systran's Hugging Face account lists no turbo model, and the library fetches turbo from mobiuslabsgmbh. Systran/faster-whisper-large-v3 was downloaded about 1.31 million times in the last 30 days, per Hugging Face's API on 26 Oct 2026.
| Name you pass | Hugging Face repo it downloads |
|---|---|
| tiny, tiny.en, base, base.en, small, small.en, medium, medium.en | Systran/faster-whisper-<name> |
| large-v1, large-v2 | Systran/faster-whisper-large-v1, Systran/faster-whisper-large-v2 |
| large-v3, large | Systran/faster-whisper-large-v3 |
| turbo, large-v3-turbo | mobiuslabsgmbh/faster-whisper-large-v3-turbo |
| distil-large-v3 | Systran/faster-distil-whisper-large-v3 |
| distil-large-v3.5 | distil-whisper/distil-large-v3.5-ct2 |
Sources:
- faster_whisper/utils.py · model name map, fetched 26 Oct 2026
- Hugging Face · Systran models (Hub API), fetched 26 Oct 2026
faster-whisper vs Whisper vs WhisperX, and what faster-whisper-xxl is
The README times 13 minutes of audio with large-v2 and beam size 5 on an RTX 3070 Ti 8 GB (CUDA 12.4). These are the maintainers' numbers, not ours:
WhisperX is not a rival engine. It runs faster-whisper underneath and adds wav2vec2 word alignment and pyannote speaker labels; its README claims 70x real time with batched large-v2 in under 8 GB of GPU memory. Use faster-whisper for transcripts with segment times, WhisperX when you need exact word timing or speakers.
Faster-Whisper-XXL is a third thing: Purfview's standalone executables of faster-whisper for Windows and Linux (whisper-standalone-win), with extra VAD methods and speaker diarization, for people who do not want Python. The Pro build goes to donors only.
| Implementation | Precision | Time | VRAM |
|---|---|---|---|
| openai/whisper | fp16 | 2m23s | 4,708 MB |
| whisper.cpp (Flash Attention) | fp16 | 1m05s | 4,127 MB |
| faster-whisper | fp16 | 1m03s | 4,525 MB |
| faster-whisper, batch_size=8 | fp16 | 17s | 6,090 MB |
| faster-whisper | int8 | 59s | 2,926 MB |
Sources:
- SYSTRAN/faster-whisper · README (GitHub), fetched 26 Oct 2026
- m-bain/whisperX · README (GitHub), fetched 26 Oct 2026
- Purfview/whisper-standalone-win · README and releases, fetched 26 Oct 2026
MIT-licensed reimplementation of OpenAI Whisper on the CTranslate2 engine. ~4× faster on a single GPU and roughly 2× less VRAM at the same accuracy — int8 on GPU lands large-v2 inference in ~59s for a 13-min clip at ~2.9 GB VRAM versus reference Whisper's 2m23s at 4.7 GB. Supports word timestamps, Silero VAD filter, and a BatchedInferencePipeline for parallel chunk decoding.
Best for production batch pipelines, single-GPU servers, anyone hitting OOM on stock Whisper, and CPU-only laptops via int8. No diarization — pair with pyannote, or step up to WhisperX. License: MIT (and CTranslate2 itself is MIT). Latest stable: 1.2.1 on PyPI, Python ≥3.9.
What it is
faster-whisper wraps Whisper in CTranslate2 — a tuned inference engine for transformer models. On a single consumer GPU it's ~4× faster than reference Whisper and uses ~2× less VRAM, with essentially identical accuracy. This is what most production Whisper stacks actually run, including Whipscribe. MIT-licensed and stable.
Watch out for: No diarization (pair with pyannote or whisperX); model conversion step can trip people up the first time.
Install / use
pip install faster-whisperSetting faster-whisper up or fixing an error? faster-whisper setup guide and common fixes.
Pick a model · 6 variants
faster-whisper users almost always pick a model before they pick a deployment. Each card below links the canonical Hugging Face model card — the library fetches the CTranslate2 weights from there on first run, no manual download. VRAM figures are at fp16 unless noted; int8 roughly halves them.
The reference Whisper-large-v3 conversion. Lowest WER in the family on most languages and the default starting point for accuracy-sensitive batch jobs. Drop-in: WhisperModel("large-v3").
~3.1 GB on disk · ~3 GB VRAM at fp16 · ~1.5 GB VRAM at int8 · 99 languages
CT2 conversion of OpenAI's whisper-large-v3-turbo (4 decoder layers instead of 32). On the same GPU you get materially higher throughput at quality close to large-v3 — the sweet spot for high-volume batch transcription.
~1.5 GB on disk · ~2 GB VRAM at fp16 · 100 languages · community CT2 conversion
HuggingFace's distilled Whisper-large-v3 in CT2 form. Roughly 6× faster than large-v3 at near-parity English WER. Trained primarily on English so accuracy on other languages will drop — use one of the larger models if you need multilingual.
~1.5 GB on disk · ~1.6 GB VRAM at fp16 · English-focused (multilingual support inherited but weaker)
Five decoder layers fewer than large, materially smaller VRAM footprint, WER still strong for non-adversarial audio. The default pick on a 6–8 GB consumer GPU when large-v3 won't fit alongside the rest of your stack.
~1.5 GB on disk · ~1.6 GB VRAM at fp16 · ~800 MB at int8 · 99 languages
What faster-whisper's own README benchmarks on CPU. With compute_type="int8" a 13-min clip runs in ~2m37s on CPU vs reference Whisper's 6m58s — usable for desk-side batch work without a GPU.
~480 MB on disk · ~1 GB RAM at int8 on CPU · 99 languages
Smallest CT2 Whisper variant. WER trails the larger models materially but the decode loop is small enough to run usable inference on edge ARM boxes or alongside heavier workloads on shared hardware.
~75 MB on disk · ~400 MB RAM at int8 · 99 languages
WhisperModel(...) and faster-whisper will download + cache on first use. For comparison, see openai/whisper (reference implementation) and WhisperX (faster-whisper plus pyannote diarization and forced alignment).Setup recipes · pick one and copy
Three runnable configurations covering the most common faster-whisper deployments. Each block is copy-and-run against faster-whisper 1.2.x.
pip install · CPU or GPU · 8-line transcription script. The default choice.
# faster-whisper 1.2.x · Python >=3.9
pip install faster-whisper
# transcribe.py
from faster_whisper import WhisperModel
# device="cpu" + compute_type="int8" also works
model = WhisperModel(
"large-v3-turbo", # or large-v3, medium, small, tiny
device="cuda",
compute_type="float16",
)
segments, info = model.transcribe(
"audio.mp3",
beam_size=5,
)
print(f"detected language: {info.language} (p={info.language_probability:.2f})")
for s in segments:
print(f"[{s.start:6.2f} -> {s.end:6.2f}] {s.text}")CUDA + cuDNN · float16 weights · BatchedInferencePipeline for parallel chunks. This is the production sweet spot.
# CUDA + cuDNN already installed on the box.
# For CUDA 12, cuDNN 9 is bundled via nvidia-cudnn-cu12.
pip install faster-whisper nvidia-cudnn-cu12
from faster_whisper import WhisperModel, BatchedInferencePipeline
model = WhisperModel(
"large-v3-turbo",
device="cuda",
compute_type="float16", # int8_float16 halves VRAM
)
# batched = parallel decode across 30s chunks
batched = BatchedInferencePipeline(model=model)
segments, info = batched.transcribe(
"podcast.wav",
batch_size=16, # 8 on 8 GB, 16 on 16 GB, 32 on 24 GB+
beam_size=5,
)
for s in segments:
print(s.start, s.end, s.text)word_timestamps=True for force-aligned words · vad_filter=True drops silence before decode (Silero VAD).
from faster_whisper import WhisperModel
model = WhisperModel("large-v3", device="cuda", compute_type="float16")
segments, info = model.transcribe(
"meeting.wav",
beam_size=5,
word_timestamps=True,
vad_filter=True,
vad_parameters=dict(
min_silence_duration_ms=500, # drop silences >=500ms
),
)
for s in segments:
print(f"[{s.start:6.2f} -> {s.end:6.2f}] {s.text}")
for w in s.words:
print(f" {w.start:6.2f}-{w.end:6.2f} {w.word!r}")What people actually do with faster-whisper-style transcription
The tool is the means. These are the jobs — each one priced at published rates, each one wired up on its own page.
Features
| Speaker diarization | No |
| Word-level timestamps | Yes |
| Streaming / real-time | No |
| Languages supported | 99 |
| HIPAA eligible | No |
Links
- SYSTRAN/faster-whisper ↗ ↗main repo · README has the canonical benchmark table and BatchedInferencePipeline docs
- PyPI · faster-whisper 1.2.1 ↗ ↗latest stable, released 2025-10-31 · Python ≥3.9
- huggingface.co/Systran ↗ ↗SYSTRAN-maintained model collection · large-v3, large-v2, medium, small, tiny, base
- CTranslate2 docs ↗ ↗the OpenNMT inference engine faster-whisper is built on · MIT-licensed
- CTranslate2 quantization guide ↗ ↗int8 / int8_float16 / float16 / float32 tradeoffs · what compute_type to pick
- m-bain/whisperX ↗ ↗faster-whisper + pyannote diarization + forced alignment for word-accurate timestamps
- jianfch/stable-ts ↗ ↗Whisper output post-processing for more reliable timestamps and SRT/VTT/ASS export
- snakers4/silero-vad ↗ ↗the VAD model faster-whisper bundles for vad_filter=True
faster-whisper vs Whipscribe
| Feature | faster-whisper | Whipscribe |
|---|---|---|
| Category | Open source | Transcription APIs |
| Pricing | free | $4–$24 one-time packs (credits never expire) · free preview of the opening minutes |
| Speaker identification | No | Named in the summary, key quotes and chapters |
| Word timestamps | Yes | Yes |
| Streaming | No | No |
| Languages | 99 | 99 |
| Platforms | Linux, macOS, Windows, GPU | Web, API, MCP |
Alternatives to faster-whisper
Frequently asked about faster-whisper
What is the faster-whisper large-v3-turbo model called?
Pass "turbo" or "large-v3-turbo" to WhisperModel. It downloads mobiuslabsgmbh/faster-whisper-large-v3-turbo; Systran publishes no turbo repo.
Does faster-whisper need FFmpeg?
No. It decodes audio with PyAV, which bundles the FFmpeg libraries. openai-whisper, by contrast, needs the ffmpeg command installed.
What is faster-whisper-xxl?
Standalone executables of faster-whisper by Purfview (github.com/Purfview/whisper-standalone-win), with extra VAD and diarization options. It is a separate project from SYSTRAN/faster-whisper.
faster-whisper vs WhisperX: which should I use?
faster-whisper for fast transcripts. WhisperX runs faster-whisper and adds word alignment and speaker labels, at the cost of extra models and a Hugging Face token for diarization.
Is faster-whisper more accurate than OpenAI Whisper?
Accuracy is essentially identical — same model weights, same training. The difference is runtime speed and VRAM usage, both of which faster-whisper improves materially. Any measurable WER gap is in the noise.
How much faster is faster-whisper?
Published benchmarks show roughly 4x faster inference on a single GPU versus the reference openai-whisper package, with ~2x lower VRAM. Your mileage varies with batch size, model size, and hardware.
Does faster-whisper support diarization?
No. It handles transcription only. Combine it with pyannote for speaker labels, or use whisperX, which integrates both.
What license is faster-whisper?
MIT — permissive for commercial and non-commercial use. CTranslate2 (the underlying engine) is also MIT-licensed.
Can faster-whisper run on CPU?
Yes. Use compute_type='int8' and a smaller model (base or small) for acceptable speed. For production CPU workloads, whisper.cpp is usually faster.
Does faster-whisper work on Mac, Windows and Linux?
faster-whisper runs on Linux, macOS, Windows and GPU.
How do I install faster-whisper?
With pip: `pip install faster-whisper`. You need Python and, for most audio tools, ffmpeg on your PATH first.
How many languages does faster-whisper support?
faster-whisper lists 99 languages. That is the Whisper-family multilingual set, so quality varies by language — English and the large European languages are strongest.
Does faster-whisper give word-level timestamps?
Yes — faster-whisper produces timing per word, which is what subtitle cues and clip boundaries need.
Can faster-whisper transcribe live audio?
No — faster-whisper works on finished files, not a live stream.
Is faster-whisper HIPAA compliant?
faster-whisper does not claim HIPAA compliance. For PHI, run it on infrastructure you control or choose a service that will sign a BAA.
Whipscribe is a hosted transcription service. If you want transcripts without running infrastructure, paste a URL or drop a file in the form below — you'll have a transcript in seconds.
What we measured, last 30 days
2,547 recordings, 1,244 audio hours, 66 languages detected. An hour of audio came back in 3.5 minutes for half of the recordings and within 18.0 minutes for nine in ten, measured from the moment the file landed to the finished transcript. Refreshed 2026-10-07; method on how we test.
Transcribe your recordings
Running faster-whisper yourself? Send the same file here and compare the transcript. Transcribe any language, any recording · Transcription services from $0.29 an hour · Transcribe from any meeting or recording platform · How to transcribe, step by step
Explore
All transcription tools · By language · By platform · By file format · How-to guides · Transcription services · Transcription API · Pricing
Related
- Whisper's own translate task writes English text only; to keep the original transcript and get English subtitles on the same timestamps, see translate Italian audio to English.
- Segments and word timestamps as a file your code can read: audio to JSON.
- Feeding a batch job? Convert to 16 kHz mono WAV first, in the browser, so every file has the same format.
- Choosing between implementations? Which open-source Whisper to use.
Have the recording already?
Skip the setup and send the file or the link: see MP3 to text for long recordings, how to transcribe a podcast and the SRT file generator.
Rather have it done for you?
Free tools cost setup time. When you want a finished transcript instead: see our transcription service with prices and turnaround.