Cohere Transcribe (03-2026)

cohere-transcribe-03-2026 is Cohere Labs' first audio model, a dedicated 2B-parameter audio-in / text-out automatic speech recognition model trained from scratch (no Whisper distillation) with supervised cross-entropy.

• Encoder-decoder cross-attention transformer with a Fast-Conformer encoder (holding the majority of the 2B parameters) and a lightweight Transformer decoder.
• Input: raw waveform → log-mel spectrogram (auto-resampled to 16 kHz, multi-channel averaged to mono).
• 16k multilingual BPE tokenizer with byte fallback.
• Customizable punctuation via prompt; automatic long-form chunking.

• ~0.5M hours of curated audio-transcript pairs + synthetic augmentations.
• Non-speech background-noise augmentation (0–30 dB SNR).
• Proprietary mix-balancing and audio decontamination checks for test/train overlap.

English, German, French, Italian, Spanish, Portuguese, Greek, Dutch, Polish, Arabic, Vietnamese, Chinese (Mandarin), Japanese, Korean. No automatic language detection — language must be specified explicitly.

• #1 on the Hugging Face Open ASR Leaderboard as of 2026-03-26 with an average WER of 5.42 across 8 English benchmarks.
• ~525× real-time (RTFx ≈ 524.88), roughly 3× faster than comparably sized ASR models.
• 61% average human preference win rate vs. competing ASR models.
• Ranks 2nd among open-source models on the multilingual ASR leaderboard.

Supported natively in transformers (CohereAsrForConditionalGeneration), vLLM (/v1/audio/transcriptions), Apple Silicon, browser, and mobile; 18 quantized variants on the Hub.