WhisperX wraps Whisper Large v3 with faster-whisper batched inference, wav2vec2 forced alignment for word-level timestamps, and pyannote diarization for speaker identification — achieving up to 70x real-time on an RTX 4090. It's the recommended pipeline for subtitle generation, podcast transcription, and meeting recording analysis. Trade-off: more dependencies and a slightly more complex install than vanilla whisper.cpp.
WhisperX needs at minimum 0.9 GB of VRAM at Q4_K_M quantization (3.4 GB at FP16). The cheapest GPU that comfortably fits with KV-cache headroom is the Intel Arc A380 6GB (6 GB VRAM, $139 MSRP). Measured throughput hits 210 x realtime on NVIDIA B200 192GB. This model fits a single consumer GPU under 48 GB, so a one-card build works.
Weights only. Add ~20-30% for KV cache at typical context lengths.
| Quant | VRAM | Quality | Notes |
|---|---|---|---|
| FP16 | 3.4 GB | Reference | Training-precision weights |
| Q8_0 | 1.7 GB | Near-lossless | 8-bit, ~0.1% perplexity hit |
| Q6_K | 1.4 GB | Very high | 6-bit, near Q8 quality |
| Q5_K_M | 1.2 GB | High | Strong middle ground |
| Q4_K_M | 0.9 GB | Balanced (recommended) | Default for local deployments |
| Q4_0 | 1.0 GB | Legacy 4-bit | Older GGUF, kept for compatibility |
| Q3_K_M | 0.7 GB | Lossy | When VRAM is very tight |
| Q2_K | 0.5 GB | Extreme | Rescue option, quality degrades visibly |
Filtered from MyAIHardware's GPU database. FP16 needs 3.4 GB; Q4_K_M needs 0.9 GB.
33 measurement(s) for this model from MyAIHardware's benchmark database.
| Device | Speed | Quant | Context | Power |
|---|---|---|---|---|
| NVIDIA B200 192GB | 210x RT | FP16 | 0 | 1000W |
| NVIDIA H100 SXM5 80GB | 145x RT | FP16 | 0 | 700W |
| NVIDIA GeForce RTX 5090 32GB | 128x RT | FP16 | 0 | 575W |
| NVIDIA GeForce RTX 5090 32GB | 105x RT | FP16 | 0 | 575W |
| NVIDIA GeForce RTX 5080 16GB | 95x RT | FP16 | 0 | 360W |
| AMD Instinct MI300X 192GB | 95x RT | FP16 | 0 | 750W |
| NVIDIA GeForce RTX 4080 Super 16GB | 82x RT | FP16 | 0 | 320W |
| NVIDIA GeForce RTX 4090 24GB | 78x RT | FP16 | 0 | 450W |
| NVIDIA GeForce RTX 5090 32GB | 78x RT | FP16 | 0 | 575W |
| NVIDIA GeForce RTX 4080 Super 16GB | 64x RT | FP16 | 0 | 320W |
| NVIDIA GeForce RTX 3090 24GB | 56x RT | FP16 | 0 | 350W |
| AMD Radeon RX 7900 XTX 24GB | 48x RT | FP16 | 0 | 355W |
| NVIDIA GeForce RTX 4080 Super 16GB | 48x RT | FP16 | 0 | 320W |
| NVIDIA GeForce RTX 4070 12GB | 42x RT | FP16 | 0 | 200W |
| Apple M4 Max (40c GPU, 128GB) | 32x RT | FP16 | 0 | 70W |
| NVIDIA GeForce RTX 3090 24GB | 32x RT | FP16 | 0 | 350W |
| NVIDIA GeForce RTX 4060 Ti 16GB | 30x RT | FP16 | 0 | 165W |
| Intel Arc B580 12GB | 28x RT | INT8 | 0 | 190W |
| Apple M3 Max (40c GPU, 64GB) | 26x RT | FP16 | 0 | 65W |
| Apple M4 Pro (20c GPU, 48GB) | 22x RT | FP16 | 0 | 50W |
Official weights + popular runtime tags.
Step-by-step guides that use this model.
Models in a similar size or capability class.
Use these in papers, blog posts, or internal docs.
@misc{whisperx-2023,
title={WhisperX},
author={OpenAI},
year={2023},
url={https://github.com/m-bain/whisperX}
}OpenAI (2023). WhisperX [Model card]. Hugging Face. https://github.com/m-bain/whisperX
WhisperX (Whisper, OpenAI, 2023). Available at https://github.com/m-bain/whisperX.