README.md

---
license: apache-2.0
language:
- en
tags:
- asr
- speech-to-text
- benchmark
- rtx-5090
- parakeet
- whisper
- local-ai
pretty_name: Sovereign ASR Bench (RTX 5090)
configs:
- config_name: default
  data_files: board.csv
---

# Sovereign ASR Bench — RTX 5090

Local, self-hosted **automatic speech recognition** benchmarks on one RTX 5090 32GB.
Part of the WITCHEER local-AI rig. Methodology that matters: **load-once** measurement (so
RTFx times transcription, not model load), **one shared text normalizer** applied to every
model output *and* reference, and **micro-averaged WER** (total errors / total reference
words — the LibriSpeech standard). The board lives as data in `board.csv` (shown in the viewer).

## Board — LibriSpeech (test-clean / test-other), 200 utts/split, load-once

| model | params | arch | WER clean | WER other | RTFx | VRAM |
|---|---|---|---|---|---|---|
| **parakeet-tdt-0.6b-v2** | 0.6B | TDT transducer | 1.49% | **4.73%** | **451×** | **2.0 GB** |
| whisper-large-v3 | 1.55B | attention enc-dec | 1.47% | 5.96% | 53× | 4.6 GB |
| whisper-large-v3-turbo | 809M | attention enc-dec | 1.42% | 6.26% | 117× | 2.5 GB |

![WER-other vs RTFx vs VRAM](./assets/asr-head-to-head.png)

Clean read speech is a three-way tie (saturated). On the **noisy** split the 0.6B Parakeet
**transducer** wins on WER **and** runs 4–9× faster on the least VRAM — the smallest model is
the most noise-robust. Full writeup + mechanism in [`asr-head-to-head.md`](./asr-head-to-head.md).

## Method

- **Runner (load-once):** [`scripts/asr_bench.py`](https://github.com/notwitcheer/llm-bench-rig) —
  parakeet.cpp `bench --manifest` (TDT) + whisper.cpp multi-file, both on sm_120 CUDA.
- **Scorer (dep-free):** `scripts/asr_report.py` — micro-WER + RTFx + peak VRAM.

## Caveats

200 utts/split is **directional** (ranking solid, exact margin provisional vs the full 2,939-utt
test-other). **English-only.** Number-words-vs-digits are **not** reconciled — counted as errors
equally for every model.