cstr/omniasr-llm-unlimited-300m-v2-GGUF

OmniASR-LLM-Unlimited-300M-v2 — GGUF

GGUF conversion of Meta's OmniASR-LLM-Unlimited-300M-v2 from facebookresearch/omnilingual-asr for the CrispASR omniasr backend.

1600+ language ASR with a 300M encoder + LLaMA decoder (~1.6B total params). The "Unlimited" variant handles arbitrarily long audio via a 15-second sliding-segment protocol.

Architecture

• —Input: raw 16 kHz mono PCM (no mel features)
• —Frontend: 7-layer CNN positional conv stack
• —Encoder: 24-layer transformer (1024-dim, 4096 FFN, 16 heads), vocab=9812
• —Projection: enc_proj bridge to decoder
• —Decoder: 12-layer LLaMA (4096-dim, 2816 FFN, 8 heads)
• —Tokenizer: SentencePiece; output: UTF-8 text
• —Unlimited mode: 15s-segment streaming for audio of any length

Files

Usage

crispasr --backend omniasr -m omniasr-llm-unlimited-300m-v2-q4_k.gguf \
    -f audio.wav -l auto

License

Apache 2.0.

Credits

• —Original model: Meta / facebookresearch
• —GGUF conversion & runtime: CrispASR

Provenance and EU AI Act Art. 53 note

• —Upstream model: facebookresearch/omnilingual-asr (Meta OmniASR-LLM-Unlimited-300M-v2).
• —Upstream licence: apache-2.0. This repository redistributes under the same terms; it grants no rights the upstream licence does not.
• —What was done here: format conversion and/or quantisation only (GGUF). No training, no fine-tuning, no merging, no distillation, no change to architecture, vocabulary or capability. Only the numeric representation of the upstream weights differs.
• —Training data: documented — where it is documented at all — by the upstream provider; see the upstream model card. No training data was used, added or selected by this repository.
• —Provider status: under Regulation (EU) 2024/1689 the upstream authors remain the provider of this model. Converting the serialisation format does not make this repository the provider of a new general-purpose AI model, and no such claim is made. Questions about training content, copyright policy or model capability belong upstream.