TDDE19 · FASTTALK · DEC 2025
Three models, one voice loop.
Whisper large-v3-turbo · Kokoro TTS · Q4 · all local
Qwen3 8B
Success rate93.3%
User-perceived latency, median280 ms
Llama 3.1 8B
Success rate86.7%
User-perceived latency, median308 ms
Ministral 3 8B
Success rate100%
User-perceived latency, median320 ms
····
Latency per run
User-perceived latency in ms, from the end of speech to the first audio back.
Qwen3 8BLlama 3.1 8BMinistral 3 8B
What each model is best at
Qwen3 8B
- –Fastest median UPL (280 ms)
- –Fastest median time to start reply (277 ms)
- –Lowest speech-to-text spread (σ 27 ms)
- –Best on short questions
But: one 929 ms outlier
Llama 3.1 8B
- –Lowest UPL spread (σ 27 ms)
- –Tightest P5 to P95 range (69 ms)
- –No extreme outliers (max 378 ms)
- –Most concise answers
But: 4 failed long-question runs
Ministral 3 8B
- –100 % success (30 of 30)
- –No speech-to-text failures
- –Most detailed answers
- –Handles long questions well
But: slowest overall (320 ms median)
Which to use
All three stay under 440 ms P95 for conversational turns, well inside the one-second target. Llama 3.1 8B gives the best experience because its latency is predictable. Qwen3 8B is the choice when raw speed matters more than consistency, and Ministral 3 8B when every run has to succeed.