whatmodel.sh · find the best speech-to-text models you can run

Best speech-to-text models for long recordings to run locally on a 16GB graphics card

A leaderboard of the best speech-to-text models for long recordings you can run locally on a 16GB graphics card (with 32GB of system RAM). Cohere Transcribe (03-2026) is #1 right now, followed by Canary-Qwen-2.5B and Whisper large-v3. Models are ranked on transcription tests with long recordings. Each model shows the download that fits and the command to run it.

25 models · scores combine 1 benchmarks · updated how it works · find a leaderboard · benchmark your machine
task[all tasks][long-form transcription][multilingual transcription]

Scores use reviewed long-form transcription evaluations only. Compare scores within this task. Models without task measurements are excluded.

Models for 32GB RAM · 16GB graphics card → ~16GB fast · local only · reset
+ more
card[RTX 4060 Ti 16GB][RTX 4070 Ti Super][RTX 4080][RTX 4080 Super][RTX 5060 Ti 16GB][RTX 5070 Ti][RTX 5080][RTX 3080 Laptop][RTX 3080 Ti Laptop][RTX 4090 Laptop][RTX 5080 Laptop][RTX A4000][RX 7600 XT][RX 7800 XT][RX 7900 GRE][RX 9060 XT][RX 9070][RX 9070 XT]
set[all sets][Open ASR English long-form WER · 31][Voxtral report · English long-form 4-set WER · 4]

Filter the model list while keeping task scores. Models without a shared score scale are not yet calibrated; select a set to see its own scores.

Detect your exact chip or graphics card:

curl -G whatmodel.sh/stt --data-urlencode "mac=$(sysctl -n hw.memsize machdep.cpu.brand_string)"Mac

curl -G whatmodel.sh/stt --data-urlencode "ram=$(awk '/MemT/{print $2}' /proc/meminfo)" --data-urlencode "nv=$(nvidia-smi --query-gpu=name,memory.total --format=csv,noheader 2>/dev/null)"Linux

curl.exe -G whatmodel.sh/stt --data-urlencode "ram=$((gcim Win32_ComputerSystem).TotalPhysicalMemory)" --data-urlencode "nv=$(nvidia-smi --query-gpu=name,memory.total --format=csv,noheader 2>$null)"Windows

$ terminal

detecting your OS…

Run curl "whatmodel.sh/stt?ram=32&gpu=card&vram=16&max=0"to get this list in your terminal (add &format=json for agents).

ram[8][16][24][32][48][64][96][128][192][256]
[]
GB
gpu[apple silicon][graphics card][unified memory][none]
vram[4][8][12][16][24][32][48][64][80][96][128]
[]
GB
price[local][$0.001][$0.002][$0.005][any]max per audio min

detecting your OS…

Score vs memory to run

#modeltask score↓where↓speed↓RAM↓get itsources
#modeltask score↓where↓speed↓RAM↓get itsources
5Cohere Transcribe (03-2026) · bf16 Cohere63±10*local~5.5GBhow to run › Open ASR Leaderboard
7Canary-Qwen-2.5B · bf16 NVIDIA59±10*local~6.6GBhow to run › Open ASR Leaderboard
8Whisper large-v3 · fp16 OpenAI58±10*local~4.2GBhow to run › Open ASR Leaderboard
9Whisper large-v3-turbo · fp16 OpenAI58±10*local~2.4GBhow to run › Open ASR Leaderboard
11Parakeet TDT 0.6B v2 · fp32 NVIDIA56±10*local~3.4GBhow to run › Open ASR Leaderboard
12Parakeet TDT 0.6B v3 · fp32 NVIDIA55±10*local~3.5GBhow to run › Open ASR Leaderboard
13Parakeet CTC 1.1B · fp32 NVIDIA54±10*local~5.6GBhow to run › Open ASR Leaderboard
14Whisper large (v1) · fp32 OpenAI54±10*local~7.9GBhow to run › Open ASR Leaderboard
15Whisper large-v2 · fp32 OpenAI54±10*local~7.9GBhow to run › Open ASR Leaderboard
16Parakeet RNNT 0.6B · fp32 NVIDIA53±10*local~3.5GBhow to run › Open ASR Leaderboard
18Distil-Whisper distil-large-v3.5 · fp16 Hugging Face (Distil-Whisper)52±10*local~2.3GBhow to run › Open ASR Leaderboard
19Parakeet CTC 0.6B · fp32 NVIDIA52±10*local~3.4GBhow to run › Open ASR Leaderboard
20Distil-Whisper distil-large-v3 · fp16 Hugging Face (Distil-Whisper)51±10*local~2.3GBhow to run › Open ASR Leaderboard
21Parakeet TDT 1.1B · fp32 NVIDIA50±10*local~5.8GBhow to run › Open ASR Leaderboard
22Parakeet RNNT 1.1B · fp32 NVIDIA47±10*local~5.6GBhow to run › Open ASR Leaderboard
23STT En Conformer-CTC Large · fp32 NVIDIA45±10*local~1.1GBhow to run › Open ASR Leaderboard
25STT En Conformer-Transducer Small · fp32 NVIDIA34±10*local~0.6GBhow to run › Open ASR Leaderboard
26Distil-Whisper distil-large-v2 · fp16 Hugging Face (Distil-Whisper)30±10*local~2.3GBhow to run › Open ASR Leaderboard
27STT En Conformer-CTC Small · fp32 NVIDIA28±10*local~0.6GBhow to run › Open ASR Leaderboard
28STT En Conformer-Transducer Large · fp32 NVIDIA27±10*local~1.1GBhow to run › Open ASR Leaderboard
29STT En FastConformer-CTC Large · fp32 NVIDIA27±10*local~1.1GBhow to run › Open ASR Leaderboard
30STT En FastConformer-Transducer Large · fp32 NVIDIA24±10*local~1.1GBhow to run › Open ASR Leaderboard
31Distil-Whisper distil-medium.en · fp16 Hugging Face (Distil-Whisper)16±10*local~1.4GBhow to run › Open ASR Leaderboard
—Kyutai STT 2.6B (en) · bf16 Kyutainot yet calibratedlocal~6.7GBhow to run › Voxtral Realtime
—Voxtral Mini 4B Realtime (2602) · bf16 Mistralnot yet calibratedlocal~11GBhow to run › Voxtral Realtime

+ 10 more don't fit your machine or budget · show them

35 of 35 catalogue models listed here · sources and licences

* new or thinly tested: ranked on too few benchmark results · † only the maker’s own numbers so far · ~RAM is estimated; full context may need more memory