Parakeet TDT 0.6B v3 vs Cohere Transcribe: which should you use?
Parakeet v3 covers 25 European languages with timestamps under CC BY 4.0; Cohere Transcribe covers 14 languages including Chinese, Japanese and Korean under Apache 2.0.
Verdict. Choose by languages first, then by whether you need timestamps. Parakeet TDT 0.6B v3 transcribes 25 European languages, detects the language itself, adds punctuation and word-level timestamps, runs in NVIDIA's NeMo or in Transformers, and handles up to 24 minutes of audio in one pass on an 80 GB GPU. It has no Chinese, Japanese, Korean or Arabic, and its CC BY 4.0 license means you must credit NVIDIA. Cohere Transcribe (2.07B) covers 14 languages including Mandarin Chinese, Japanese, Korean and Arabic under Apache 2.0. Cohere reports an average word error rate of 5.42% on the English Open ASR Leaderboard in March 2026, and NVIDIA reports 6.34% for Parakeet on the same group of English datasets. The two were measured by their own developers at different times, so treat that as a rough guide. Cohere Transcribe has no language detection, timestamps or speaker labels, and Hugging Face asks for your contact details before it releases the weights.
Specifications
| Specifications | Parakeet TDT 0.6B v3 | Cohere Transcribe |
|---|---|---|
| Developer | NVIDIA | Cohere |
| Released | August 14, 2025 | March 26, 2026 |
| Total parameters | 627M (Parakeet TDT 0.6B v3) | 2.07B (Cohere Transcribe 03-2026) |
| Active parameters | All | All |
| Architecture | Other | Other |
| Context window | – | – |
| Input | Audio | Audio |
| License | CC BY 4.0 (Permissive) | Apache 2.0 (Permissive) |
| Memory for the smallest variant (4-bit) | 2.89 GB (Parakeet TDT 0.6B v3) | 4.75 GB (Cohere Transcribe 03-2026) |
Performance claims come from the developer and have not been independently verified.
Pick Parakeet TDT 0.6B v3 if
- You transcribe European languages and want the language detected automatically.
- You need word-level timestamps for subtitles or search.
- You want a smaller 0.6B model, with a 0.71 GB 8-bit GGUF file for NeMo-Speech.cpp.
- You can credit NVIDIA under CC BY 4.0 and run on Linux with NVIDIA GPUs.
Pick Cohere Transcribe if
- You need Chinese, Japanese, Korean or Arabic, which Parakeet v3 does not cover.
- You want Apache 2.0, with no credit requirement.
- You know the language in advance and care most about English accuracy on the leaderboard average.
- You transcribe long recordings in batches, since Cohere says offline throughput is up to three times that of similarly sized dedicated recognizers.