Approximate accuracy comparison between locally hosted Whisper models of varying sizes (tiny, small, medium, large), based on their typical Word Error Rates.

local-whisper-accuracy-estimates.md

Q&A: How Much Accuracy Do You Lose When Using Smaller Whisper Models?

This is a summary of a prompt in which the user asked:

"If I run Whisper locally and choose between tiny, small, medium, and large models, how much accuracy difference should I expect between them?"

The goal was to understand how much less accurate smaller Whisper models are compared to the larger ones.

---

Accuracy Comparison (Approximate)

Model Size	Relative Accuracy	Estimated Word Error Rate (WER)	Notes
Large	100% (baseline)	Lowest WER (~4-5%)	Most accurate, but computationally heavy
Medium	~99%	~5-6%	Nearly as good as large, much faster
Small	~97–98%	~6-8%	Good speed/accuracy balance for most cases
Tiny	~90–95%	~10–15%+	Fastest, but significantly less accurate

---

Summary of Accuracy Differences

• Medium vs. Large: ~0.5–1% less accurate
• Small vs. Medium: ~1–2% less accurate
• Tiny vs. Small: ~5–8% less accurate

These figures are approximations and vary depending on:

• Language
• Audio clarity
• Domain vocabulary
• Noise levels in recordings

---

This gist was generated with the help of OpenAI based on the user's prompt and publicly available model benchmarks (as of May 2025).