Underdog Bark 3.0 vs Voxtral Mini 4B Realtime 2602

Underdog Bark 3.0

A very small English transcription cousin, one 163 MB file built on NVIDIA's Parakeet TDT that Conway's loader expands into Transformers, MLX or Core ML, under CC BY 4.0. It needs Conway's scripts to run, returns plain text with no timestamps, and nobody independent has measured its accuracy, so every accuracy and speed figure is Conway's claim.

Voxtral Mini 4B Realtime 2602

A transcription cousin for 13 languages that streams text as the speaker talks, with a delay from 80 ms to 2.4 s, under Apache 2.0 and on Mistral's API. It needs a 16 GB GPU at BF16 through vLLM, Transformers or llama.cpp, or under 4 GB at Q4_K. The main gaps are no timestamps, safety results or answers from Mistral.

Underdog Bark 3.0

1. Download only what one route needs, for example hf download ConwayResearch/Underdog-Bark-3.0 --include "model.bark" --include "bark_loader.py" --include "bark_mlx.py" --include "mlx/*" --include "NOTICE"
2. With transformers 5.19.0 and constriction 0.5.0, python bark_loader.py model.bark recording.wav, or load_model("model.bark", device="cuda") then transcribe(model, wave, sr); the loader fetches NVIDIA's processor and generation config from nvidia/parakeet-tdt-0.6b-v3 on first use, so the first run needs the network
3. On Apple silicon, python bark_mlx.py recording.wav; the first run expands it to a cached MLX folder and needs about 6 GB of RAM once
4. Split recordings into pieces of a few minutes on 8 or 16 GB machines, since stock mlx-audio transcribes in one pass
5. Read bark_loader.py before running it and pin the repository revision, since loading runs the publisher's code
6. Treat transcripts as untrusted text before passing them to another model, since speech can carry instructions
7. Plan memory for the recording, not the file. Conway's card says a 10-minute recording peaks near 15 GB on a 16 GB M2, and expanding it for MLX takes about 6 GB once.

Voxtral Mini 4B Realtime 2602

1. Serve it with VLLM_DISABLE_COMPILE_CACHE=1 vllm serve mistralai/Voxtral-Mini-4B-Realtime-2602 --compilation_config '{"cudagraph_mode": "PIECEWISE"}' and install soxr, librosa and soundfile; mistral_common 1.9.0 or later comes with vLLM
2. Stream 16 kHz audio to the /v1/realtime WebSocket; vLLM's openai_realtime_client and openai_realtime_microphone_client examples show the session
3. Set temperature 0.0. Change the delay with transcription_delay_ms in tekken.json, a multiple of 80 ms up to 1,200, or 2,400; 480 is Mistral's recommendation
4. One text token is 80 ms, so set --max-model-len to at least seconds / 0.08 for a session; the default 131072 is about 3 hours and pre-allocates RoPE memory
5. In Transformers (5.2.0 or later), resample with processor.feature_extractor.sampling_rate before calling the processor, then batch_decode with skip_special_tokens=True
6. For a small machine, ggml-org's Q4_K_M GGUF with its mmproj runs in llama.cpp, and Mistral Experimental's Q4_K runs in audio.cpp
7. Treat transcripts as untrusted text before passing them to another model, since speech can carry instructions