Meta's 7.8B omnilingual ASR model lands on Replicate with 1,693-language coverage

Meta's omnilingual ASR model lands on Replicate

A 7.8-billion-parameter automatic speech recognition model from Meta is now available to developers through Replicate, where it runs under the name meta-omnilingual-asr-7b with packaging maintained by Subformer. According to a beginner's guide published on dev.to by AImodels.fyi, the model covers 1,693 languages and posts a character error rate below 10 for 78% of them, making it Meta's suggested pick for jobs where transcription quality outweighs throughput.

Architecture and performance

The model combines a wav2vec2 encoder with a decoder based on a large language model, an approach that lets it transcribe languages zero-shot or few-shot rather than requiring paired training data for each one. The Replicate release uses the newer v2 architecture, which the guide reports as more accurate than the original LLM-ASR version. Meta trained it on more than 1.6 million hours of audio drawn from the Meta Omnilingual Audio Dataset plus supplementary multilingual corpora, with the architecture implemented in the fairseq2 toolkit.

On the hardware side it is heavyweight: the guide lists roughly 17GB of VRAM for inference at batch size one on 30-second clips in BF16 precision on an A100, along with a 30GB model download on first use. Speed is close to real time, with a real-time factor of 0.097, meaning a 30-second sample transcribes in about 2.9 seconds on that hardware.

Constraints to plan around

Several limits will shape how teams deploy it:

• Audio is capped at 40 seconds for the standard variants. A separate unlimited-length variant, omniASR_LLM_Unlimited_7B_v2, matches accuracy on longer recordings but does not currently support fine-tuning, and manually splitting audio can introduce boundary artifacts that need cleanup.
• Output is plain text only. There are no timestamps, speaker diarization, confidence scores or word-level alignments, so anything needing timing information requires a separate alignment step.
• Accuracy is uneven. The remaining 22% of supported languages degrade substantially, and some low-resource languages may yield unusable output. Per-language CER figures ship as a CSV in the README so teams can check expectations before committing.
• Languages are specified with ISO 639-3 codes plus ISO 15924 script codes, such as eng_Latn or cmn_Hans. Invalid codes fall back to automatic detection, which is slower and can misidentify similar or code-switched audio.
• The VRAM and download footprint rules out edge devices and serverless environments without GPU acceleration.

Local use outside Replicate is possible through the omnilingual-asr pip package (macOS also needs libsndfile installed), with weights cached under the fairseq2 assets directory on first run. Batch processing is supported with configurable batch sizes.

How it compares

The guide also positions the model against its neighbors on Replicate. twangodev's qwenasr specializes in Mandarin, Cantonese and related languages and may be more accurate on East Asian audio; lucataco's seamless_communication adds machine translation on top of transcription; and NexaAI's OmniAudio-2.6B targets low-latency edge deployment in a much smaller package. The Meta model stands out on breadth — a single system spanning more than 1,600 languages — rather than on Chinese-specific accuracy, translation or on-device efficiency.

Why it matters

Most commercial ASR products support a few dozen languages at best. A single hosted model that transcribes more than 1,600 of them, including hundreds the guide says were never previously supported by any commercial ASR system, changes what is practical for language documentation, archival digitization, media localization and cross-lingual research — without building custom datasets or per-language pipelines. Replicate availability means developers can test it through an API before committing GPU budget of their own. The caveats are real, though: uneven per-language accuracy, a 40-second clip limit on the standard build and bare-text output mean production adoption still calls for validation against the published error rates and, often, extra tooling around the model.