lang-id-voxlingua107-ecapa: transcribe.cpp GGUF

GGUF conversions of speechbrain/lang-id-voxlingua107-ecapa for use with transcribe.cpp.

Ported from upstream commit 0253049, pinned 2026-10-05. Validated against the SpeechBrain reference at transcribe.cpp commit aa4d0f47 on 2026-10-08.

Spoken language identification over 107 languages: SpeechBrain's ECAPA-TDNN trained on VoxLingua107. NOT a transcription model: a run returns the model's language labels ranked by probability, optionally restricted to a caller-chosen set. Takes 16 kHz mono WAV; scores up to the last 30 s of a clip.

Downloads

Top-1 accuracy on FLEURS multilingual (3,000 utterances), scored on cpu. Measured at transcribe.cpp 2fdb5c95 on 2026-10-07.

Open-set top-1 accuracy over all 107 labels, the mean over 15 FLEURS languages, on the first 5 s of each clip without silence trimming. Agreement with the SpeechBrain reference, per GGUF, is on the transcribe.cpp model page.

Usage

Build transcribe.cpp from source:

git clone git@github.com:handy-computer/transcribe.cpp.git
cd transcribe.cpp
cmake -B build && cmake --build build

Run on a 16 kHz mono WAV. This is a language identifier, not a transcription model: the CLI prints ranked language candidates.

build/bin/transcribe-cli -m lang-id-voxlingua107-ecapa-Q8_0.gguf --allow en,de,fr --top 3 input.wav
# language: de index=18 p=0.999998
#   ...

From the C API, use the LANGID role (include/transcribe/langid.h, see docs/langid.md).

If your audio isn't already 16 kHz mono WAV, convert it first:

ffmpeg -i input.mp3 -ar 16000 -ac 1 output.wav

See the transcribe.cpp model page for performance numbers, numerical validation, and reproduction steps.

License

Inherited from the base model: Apache-2.0. See the upstream model card for full terms.

---

Original Model Card

The section below is reproduced from speechbrain/lang-id-voxlingua107-ecapa at commit 0253049 for offline reference. The upstream card is the authoritative source.

VoxLingua107 ECAPA-TDNN Spoken Language Identification Model

Model description

This is a spoken language recognition model trained on the VoxLingua107 dataset using SpeechBrain. The model uses the ECAPA-TDNN architecture that has previously been used for speaker recognition. However, it uses more fully connected hidden layers after the embedding layer, and cross-entropy loss was used for training. We observed that this improved the performance of extracted utterance embeddings for downstream tasks.

The system is trained with recordings sampled at 16kHz (single channel). The code will automatically normalize your audio (i.e., resampling + mono channel selection) when calling classify_file if needed.

The model can classify a speech utterance according to the language spoken. It covers 107 different languages ( Abkhazian, Afrikaans, Amharic, Arabic, Assamese, Azerbaijani, Bashkir, Belarusian, Bulgarian, Bengali, Tibetan, Breton, Bosnian, Catalan, Cebuano, Czech, Welsh, Danish, German, Greek, English, Esperanto, Spanish, Estonian, Basque, Persian, Finnish, Faroese, French, Galician, Guarani, Gujarati, Manx, Hausa, Hawaiian, Hindi, Croatian, Haitian, Hungarian, Armenian, Interlingua, Indonesian, Icelandic, Italian, Hebrew, Japanese, Javanese, Georgian, Kazakh, Central Khmer, Kannada, Korean, Latin, Luxembourgish, Lingala, Lao, Lithuanian, Latvian, Malagasy, Maori, Macedonian, Malayalam, Mongolian, Marathi, Malay, Maltese, Burmese, Nepali, Dutch, Norwegian Nynorsk, Norwegian, Occitan, Panjabi, Polish, Pushto, Portuguese, Romanian, Russian, Sanskrit, Scots, Sindhi, Sinhala, Slovak, Slovenian, Shona, Somali, Albanian, Serbian, Sundanese, Swedish, Swahili, Tamil, Telugu, Tajik, Thai, Turkmen, Tagalog, Turkish, Tatar, Ukrainian, Urdu, Uzbek, Vietnamese, Waray, Yiddish, Yoruba, Mandarin Chinese).

Intended uses & limitations

The model has two uses:

• use 'as is' for spoken language recognition
• use as an utterance-level feature (embedding) extractor, for creating a dedicated language ID model on your own data

The model is trained on automatically collected YouTube data. For more information about the dataset, see here.

How to use

pip install git+https://github.com/speechbrain/speechbrain.git@develop

import torchaudio
from speechbrain.inference.classifiers import EncoderClassifier
language_id = EncoderClassifier.from_hparams(source="speechbrain/lang-id-voxlingua107-ecapa", savedir="tmp")
# Download Thai language sample from Omniglot and cvert to suitable form
signal = language_id.load_audio("speechbrain/lang-id-voxlingua107-ecapa/udhr_th.wav")
prediction =  language_id.classify_batch(signal)
print(prediction)
#  (tensor([[-2.8646e+01, -3.0346e+01, -2.0748e+01, -2.9562e+01, -2.2187e+01,
#         -3.2668e+01, -3.6677e+01, -3.3573e+01, -3.2545e+01, -2.4365e+01,
#         -2.4688e+01, -3.1171e+01, -2.7743e+01, -2.9918e+01, -2.4770e+01,
#         -3.2250e+01, -2.4727e+01, -2.6087e+01, -2.1870e+01, -3.2821e+01,
#         -2.2128e+01, -2.2822e+01, -3.0888e+01, -3.3564e+01, -2.9906e+01,
#         -2.2392e+01, -2.5573e+01, -2.6443e+01, -3.2429e+01, -3.2652e+01,
#         -3.0030e+01, -2.4607e+01, -2.2967e+01, -2.4396e+01, -2.8578e+01,
#         -2.5153e+01, -2.8475e+01, -2.6409e+01, -2.5230e+01, -2.7957e+01,
#         -2.6298e+01, -2.3609e+01, -2.5863e+01, -2.8225e+01, -2.7225e+01,
#         -3.0486e+01, -2.1185e+01, -2.7938e+01, -3.3155e+01, -1.9076e+01,
#         -2.9181e+01, -2.2160e+01, -1.8352e+01, -2.5866e+01, -3.3636e+01,
#         -4.2016e+00, -3.1581e+01, -3.1894e+01, -2.7834e+01, -2.5429e+01,
#         -3.2235e+01, -3.2280e+01, -2.8786e+01, -2.3366e+01, -2.6047e+01,
#         -2.2075e+01, -2.3770e+01, -2.2518e+01, -2.8101e+01, -2.5745e+01,
#         -2.6441e+01, -2.9822e+01, -2.7109e+01, -3.0225e+01, -2.4566e+01,
#         -2.9268e+01, -2.7651e+01, -3.4221e+01, -2.9026e+01, -2.6009e+01,
#         -3.1968e+01, -3.1747e+01, -2.8156e+01, -2.9025e+01, -2.7756e+01,
#         -2.8052e+01, -2.9341e+01, -2.8806e+01, -2.1636e+01, -2.3992e+01,
#         -2.3794e+01, -3.3743e+01, -2.8332e+01, -2.7465e+01, -1.5085e-02,
#         -2.9094e+01, -2.1444e+01, -2.9780e+01, -3.6046e+01, -3.7401e+01,
#         -3.0888e+01, -3.3172e+01, -1.8931e+01, -2.2679e+01, -3.0225e+01,
#         -2.4995e+01, -2.1028e+01]]), tensor([-0.0151]), tensor([94]), ['th'])
# The scores in the prediction[0] tensor can be interpreted as log-likelihoods that
# the given utterance belongs to the given language (i.e., the larger the better)
# The linear-scale likelihood can be retrieved using the following:
print(prediction[1].exp())
#  tensor([0.9850])
# The identified language ISO code is given in prediction[3]
print(prediction[3])
#  ['th: Thai']
  
# Alternatively, use the utterance embedding extractor:
emb =  language_id.encode_batch(signal)
print(emb.shape)
# torch.Size([1, 1, 256])

To perform inference on the GPU, add run_opts={"device":"cuda"} when calling the from_hparams method.

The system is trained with recordings sampled at 16kHz (single channel). The code will automatically normalize your audio (i.e., resampling + mono channel selection) when calling classify_file if needed. Make sure your input tensor is compliant with the expected sampling rate if you use encode_batch and classify_batch.

Warning: In the dataset and in the defaults of this model (see label_encoder.txt, the used ISO language code for Hebrew is obsolete (should be he instead of iw). The ISO language code for Javanese is incorrect (should be jv instead of jw). See issue #2396.

Limitations and bias

Since the model is trained on VoxLingua107, it has many limitations and biases, some of which are:

• Probably it's accuracy on smaller languages is quite limited
• Probably it works worse on female speech than male speech (because YouTube data includes much more male speech)
• Based on subjective experiments, it doesn't work well on speech with a foreign accent
• Probably it doesn't work well on children's speech and on persons with speech disorders

Training data

The model is trained on VoxLingua107.

VoxLingua107 is a speech dataset for training spoken language identification models. The dataset consists of short speech segments automatically extracted from YouTube videos and labeled according the language of the video title and description, with some post-processing steps to filter out false positives.

VoxLingua107 contains data for 107 languages. The total amount of speech in the training set is 6628 hours. The average amount of data per language is 62 hours. However, the real amount per language varies a lot. There is also a seperate development set containing 1609 speech segments from 33 languages, validated by at least two volunteers to really contain the given language.

Training procedure

See the SpeechBrain recipe.

Evaluation results

Error rate: 6.7% on the VoxLingua107 development dataset

Referencing SpeechBrain

@misc{speechbrain,
  title={{SpeechBrain}: A General-Purpose Speech Toolkit},
  author={Mirco Ravanelli and Titouan Parcollet and Peter Plantinga and Aku Rouhe and Samuele Cornell and Loren Lugosch and Cem Subakan and Nauman Dawalatabad and Abdelwahab Heba and Jianyuan Zhong and Ju-Chieh Chou and Sung-Lin Yeh and Szu-Wei Fu and Chien-Feng Liao and Elena Rastorgueva and François Grondin and William Aris and Hwidong Na and Yan Gao and Renato De Mori and Yoshua Bengio},
  year={2021},
  eprint={2106.04624},
  archivePrefix={arXiv},
  primaryClass={eess.AS},
  note={arXiv:2106.04624}
}

Referencing VoxLingua107

@inproceedings{valk2021slt,
  title={{VoxLingua107}: a Dataset for Spoken Language Recognition},
  author={J{\"o}rgen Valk and Tanel Alum{\"a}e},
  booktitle={Proc. IEEE SLT Workshop},
  year={2021},
}

About SpeechBrain

SpeechBrain is an open-source and all-in-one speech toolkit. It is designed to be simple, extremely flexible, and user-friendly. Competitive or state-of-the-art performance is obtained in various domains. Website: https://speechbrain.github.io/ GitHub: https://github.com/speechbrain/speechbrain