cstr/ecapa-lid-107-GGUF

ECAPA-TDNN Language Identification (GGUF)

GGUF conversion of speechbrain/lang-id-voxlingua107-ecapa for use with CrispASR.

Model Details

• Architecture: ECAPA-TDNN (SE-Res2Net + Attentive Statistical Pooling)
• Parameters: 21M
• Size: 43 MB (F16)
• Languages: 107 (VoxLingua107 dataset)
• License: Apache 2.0
• Training: SpeechBrain on VoxLingua107 (6,628 hours YouTube speech)

Usage with CrispASR

# As LID pre-step for any ASR backend
crispasr -m whisper-large-v3.gguf --lid-backend ecapa -l auto -f audio.wav

# Model auto-downloads on first use, or specify path:
crispasr -m model.gguf --lid-backend ecapa --lid-model ecapa-lid-107-f16.gguf -l auto -f audio.wav

Accuracy

Tested on 12-language edge-TTS benchmark (3 samples per language):

Language	Accuracy	Confidence
English	3/3	p≥0.99
German	3/3	p≥0.99
French	3/3	p≥0.99
Spanish	3/3	p≥0.96
Japanese	3/3	p≥0.99
Chinese	3/3	p≥0.99
Korean	3/3	p≥0.99
Russian	3/3	p≥0.99
Arabic	3/3	p≥0.99
Hindi	3/3	p≥0.99
Portuguese	3/3	p≥0.99
Italian	3/3	p≥0.99

Files

File	Size	Description
ecapa-lid-107-f16.gguf	43 MB	F16 weights (recommended)

Conversion

python models/convert-ecapa-tdnn-lid-to-gguf.py \
    --input speechbrain/lang-id-voxlingua107-ecapa \
    --output ecapa-lid-107-f16.gguf

Architecture

Input: 16kHz PCM → 60-dim mel fbank (SpeechBrain STFT, n_fft=400)
       → Sentence-level mean normalization
       → Block0: Conv1d(60→1024, k=5) + ReLU + BN
       → Block1-3: SE-Res2Net (1024 channels, 8 sub-bands, dilations 2/3/4)
       → MFA: concatenate block1-3 outputs → Conv1d(3072→3072, k=1) + ReLU + BN
       → ASP: Attentive Statistical Pooling → [6144]
       → BN + FC(6144→256) → embedding
       → Classifier: BN → Linear(256→512) + BN + LeakyReLU → Linear(512→107)

Citation

@inproceedings{ravanelli2021speechbrain,
  title={SpeechBrain: A General-Purpose Speech Toolkit},
  author={Ravanelli, Mirco and others},
  booktitle={Proceedings of the 22nd Annual Conference of the International Speech Communication Association (INTERSPEECH)},
  year={2021}
}

Provenance and EU AI Act Art. 53 note

• Upstream model: speechbrain/lang-id-voxlingua107-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.