Skip to content

supported_nemo_models.md

Latest commit

 

History

History
108 lines (85 loc) · 5.74 KB

supported_nemo_models.md

File metadata and controls

108 lines (85 loc) · 5.74 KB
Raw

Supported NeMo STT Models for Conversion to GGUF

This document contains a curated list of Automatic Speech Recognition (ASR) models hosted on Hugging Face that are compatible with the local C++ engine (parakeet-cli / parakeet.cpp).

To use any of these models in the app, you must:

  1. Download the .nemo file from the respective Hugging Face repository.
  2. Run the NeMo python conversion script to extract the CTC Head and convert it into a .gguf file.

Why these models?
The local C++ engine natively supports CTC (Connectionist Temporal Classification) decoding. It does not support Seq2Seq (Canary), TDT, or pure RNNT models. The models listed here are either pure Conformer-CTC models or FastConformer-Hybrid (Transducer + CTC) models. In hybrid models, the conversion script easily discards the incompatible Transducer head and preserves the compatible CTC head.

1. Official NVIDIA Models (European & Global)

NVIDIA maintains a rich library of FastConformer and Conformer models. The _pc suffix means the model automatically predicts punctuation and capitalization.

Western European

  • English:
    • nvidia/stt_en_fastconformer_hybrid_large_pc (115M params, Punctuation/Capitalization)
    • nvidia/stt_en_conformer_ctc_large (120M params, Pure CTC)
  • French:
    • nvidia/stt_fr_fastconformer_hybrid_large_pc (115M params)
    • nvidia/stt_fr_conformer_ctc_large (120M params, trained on 1500+ hours)
  • German:
    • nvidia/stt_de_fastconformer_hybrid_large_pc (115M params)
    • nvidia/stt_de_conformer_ctc_large (120M params)
  • Spanish:
    • nvidia/stt_es_fastconformer_hybrid_large_pc
  • Italian:
    • nvidia/stt_it_fastconformer_hybrid_large_pc
  • Portuguese:
    • nvidia/stt_pt_fastconformer_hybrid_large_pc
  • Dutch:
    • nvidia/stt_nl_fastconformer_hybrid_large_pc (115M params)
  • Polish:
    • nvidia/stt_pl_fastconformer_hybrid_large_pc (115M params)

Eurasian & Middle Eastern

  • Russian & Kazakh:
    • nvidia/stt_kk_ru_fastconformer_hybrid_large (Bilingual model)
  • Arabic:
    • nvidia/stt_ar_fastconformer_hybrid_large_pcd_v1.0 (Includes Diacritics)
  • Persian:
    • nvidia/stt_fa_fastconformer_hybrid_large
  • Georgian:
    • nvidia/stt_ka_fastconformer_hybrid_large_pc

2. Official AI4Bharat Models (Indic Languages)

AI4Bharat built the "MahaDhwani" and "IndicConformer" projects. All their models use the Hybrid (CTC + RNNT) architecture, making them perfectly compatible with our conversion script.

The Master Multilingual Model

  • 22 Languages (All-in-one):
    • ai4bharat/indic-conformer-600m-multilingual (600M parameters. Supports Assamese, Bengali, Bodo, Dogri, Gujarati, Hindi, Kannada, Kashmiri, Konkani, Maithili, Malayalam, Manipuri, Marathi, Nepali, Odia, Punjabi, Sanskrit, Santali, Sindhi, Tamil, Telugu, and Urdu).

Individual Language Models (Smaller & Faster)

These individual models are around 120M parameters and are much faster for specific single-language workflows.

  • Hindi: ai4bharat/indicconformer_stt_hi_hybrid_ctc_rnnt_large
  • Punjabi: ai4bharat/indicconformer_stt_pa_hybrid_ctc_rnnt_large
  • Tamil: ai4bharat/indicconformer_stt_ta_hybrid_ctc_rnnt_large
  • Gujarati: ai4bharat/indicconformer_stt_gu_hybrid_ctc_rnnt_large
  • Kannada: ai4bharat/indicconformer_stt_kn_hybrid_ctc_rnnt_large
  • Kashmiri: ai4bharat/indicconformer_stt_ks_hybrid_ctc_rnnt_large
  • Urdu: ai4bharat/indicconformer_stt_ur_hybrid_ctc_rnnt_large (AI4Bharat hosts similar repositories for the rest of the 22 Indic languages under the same naming convention).

3. Community & Independent Organizations

Because the NVIDIA NeMo toolkit is open-source, researchers around the world train custom .nemo models. Below are examples of verified community NeMo models that use the FastConformer architecture.

  • French (by Bofeng Huang):
    • bofenghuang/stt_fr_fastconformer_hybrid_large
    • Description: Trained by an independent AI researcher on over 2500 hours of open-source French data. Highly optimized for predicting hyphens and apostrophes natively.
  • French (by LINAGORA):
    • linagora/linto_stt_fr_fastconformer
    • Description: An open-source model trained by a French open-source software company, specialized for various French dialects and audio sources.

How to Find More Community Models

You can discover hundreds more by navigating to Hugging Face and searching:

  1. Under Tasks, select Automatic Speech Recognition.
  2. Under Libraries, select NeMo.
  3. Type "conformer ctc" or "fastconformer" in the search bar.

Important: Ignore models from frameworks like icefall or sherpa-onnx. Ensure the repository provides a .nemo file so the Python conversion script can successfully parse it into .gguf.

4. Additional Independent Developers & Orgs

Here are some specialized Conformer-CTC models built by the wider community:

  • Kinyarwanda (by mbazaNLP):
    • mbazaNLP/Kinyarwanda_nemo_stt_conformer_model
    • Description: An ASR model specifically tuned for Kinyarwanda, built by an African NLP research group.
  • Persian (by Ali Farokh):
    • alifarokh/nemo-conformer-medium-fa
    • Description: An independently trained, lightweight medium Conformer model for Persian.
  • Vietnamese (by Quan Hoang Ngoc):
    • QuanHoangNgoc/conformer-nemo
    • Description: A Conformer-based STT model optimized for Vietnamese.
  • Medical Dictation (by Various Orgs):
    • Search HuggingFace for tags like nemo medical or nemo-stt-medical to find models fine-tuned specifically on medical terminology and physician dictations.