cstr/ecapa-lid-commonlanguage-GGUF

ECAPA-TDNN CommonLanguage LID (GGUF)

GGUF conversion of speechbrain/lang-id-commonlanguage_ecapa for use with CrispASR.

Model Details

• Architecture: ECAPA-TDNN (SE-Res2Net + Attentive Statistical Pooling)
• Parameters: 21M
• Size: 40 MB (F16)
• Languages: 45 (CommonLanguage dataset)
• Input: 80-dim mel filterbank (vs 60 for VoxLingua107)
• Embedding dim: 192 (vs 256 for VoxLingua107)
• Classifier: Cosine similarity (vs DNN for VoxLingua107)
• License: Apache 2.0
• Training: SpeechBrain on CommonLanguage

Languages (45)

Arabic, Basque, Breton, Catalan, Chinese_China, Chinese_Hongkong, Chinese_Taiwan, Chuvash, Czech, Dhivehi, Dutch, English, Esperanto, Estonian, French, Frisian, Georgian, German, Greek, Hakha_Chin, Indonesian, Interlingua, Italian, Japanese, Kabyle, Kinyarwanda, Kyrgyz, Latvian, Maltese, Mongolian, Persian, Polish, Portuguese, Romanian, Romansh_Sursilvan, Russian, Sakha, Slovenian, Spanish, Swedish, Tamil, Tatar, Turkish, Ukranian, Welsh

Usage with CrispASR

# As LID pre-step for any ASR backend
crispasr -m model.gguf --lid-backend ecapa --lid-model ecapa-lid-commonlanguage-f16.gguf -l auto -f audio.wav

Accuracy

Tested on TTS-generated audio samples:

Language	Detected	Confidence
English	English	0.033
German	German	0.035
French	French	0.035
Spanish	Spanish	0.033
Japanese	Japanese	0.044

Note: Cosine similarity scores are naturally lower than softmax probabilities. The model ranks correctly even with low absolute confidence.

Comparison with VoxLingua107

Feature	VoxLingua107	CommonLanguage
Languages	107	45
Input mels	60	80
Embedding	256	192
Classifier	DNN	Cosine
File size	43 MB	40 MB
Labels	ISO codes (en, de, ...)	Full names (English, German, ...)

Recommendation: Use VoxLingua107 (cstr/ecapa-lid-107-GGUF) for most use cases — wider language coverage and higher confidence scores. CommonLanguage is useful if you need the specific 45-language set or prefer full language names.

Files

File	Size	Description
ecapa-lid-commonlanguage-f16.gguf	40 MB	F16 (recommended)

Technical Notes

• Quantized versions not provided — ECAPA-TDNN's small conv kernels (K=3,5) and cosine classifier are precision-sensitive
• The SpeechBrain filterbank matrix is embedded in the GGUF for exact reproduction
• Runtime uses ggml graph for the heavy SE-Res2Net forward pass with CPU-based ASP + classifier

Citation

@inproceedings{desplanques2020ecapa,
  title={ECAPA-TDNN: Emphasized Channel Attention, Propagation and Aggregation in TDNN Based Speaker Verification},
  author={Desplanques, Brecht and Thienpondt, Jenthe and Demuynck, Kris},
  booktitle={Interspeech},
  year={2020}
}

Provenance and EU AI Act Art. 53 note

• Upstream model: speechbrain/lang-id-commonlanguage_ecapa — published by speechbrain.
• 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.