Cohere Transcribe 03-2026 vs Voxtral Mini 4B Realtime 2602

Cohere Transcribe 03-2026

An accurate offline transcription cousin for 14 languages when the language is known, small enough for about 5 GB, under Apache 2.0, and also on Cohere's API. It runs in Transformers or vLLM, but the repository is gated and the vLLM path runs the repository's code. Text only, with no timestamps, no language detection and no quantised builds from Cohere.

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.

Cohere Transcribe 03-2026

1. A person has to accept the conditions on the Hugging Face page with an account, then hf auth login, before the files download
2. Offline, pip install "transformers>=5.4.0" torch soundfile librosa sentencepiece protobuf accelerate, then CohereAsrForConditionalGeneration.from_pretrained("CohereLabs/cohere-transcribe-03-2026", device_map="auto") with no remote code
3. Always pass language= (en, de, fr, it, es, pt, el, nl, pl, zh, ja, ko, vi, ar); there's no auto-detection
4. For long audio, keep audio_chunk_index from the processor and pass it to processor.decode to reassemble the chunks
5. Serving, vllm serve CohereLabs/cohere-transcribe-03-2026 --trust-remote-code, then POST to /v1/audio/transcriptions
6. Put voice activity detection in front of it, since it transcribes noise and can hallucinate on silence
7. Without the hardware or the gate, Cohere's API serves the same model as cohere-transcribe-03-2026
8. Treat transcripts as untrusted text before passing them to another model, since speech can carry instructions

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