cstr/titanet-large-GGUF

TitaNet-Large — GGUF

GGUF conversion of nvidia/speakerverification_en_titanet_large (CC-BY-4.0) for native C/C++ inference via CrispASR.

Model

Detail	Value
Architecture	TitaNet-Large — depthwise separable Conv1D encoder + ASP decoder
Parameters	23M
Embedding dim	192 (L2-normalized)
EER	0.66% on VoxCeleb1-O cleaned
Input	16 kHz mono PCM
GGUF size	~45 MB (F16 weights, F32 batch-norm)
License	CC-BY-4.0

Architecture

Preprocessor: 16kHz → 80-bin mel spectrogram (Hann window, n_fft=512, hop=160, win=400)

Encoder (Jasper-style):
  Block 0 (prolog):  DW-Conv(80, k=3) + PW-Conv(80→1024) + BN + SE + ReLU
  Block 1:           3× DW-Sep-Conv(1024, k=7)  + SE + residual + ReLU
  Block 2:           3× DW-Sep-Conv(1024, k=11) + SE + residual + ReLU
  Block 3:           3× DW-Sep-Conv(1024, k=15) + SE + residual + ReLU
  Block 4 (epilog):  DW-Conv(1024, k=1) + PW-Conv(1024→3072) + BN + SE + ReLU

Decoder:
  ASP (Attentive Statistics Pooling): 3072 → 6144
  BN + Linear: 6144 → 192
  L2-normalize

Usage with CrispASR

Speaker enrollment

# Enroll a speaker from a reference audio clip
crispasr --enroll-speaker alice \
         --speaker-db ./speakers \
         -f alice_reference.wav

Speaker identification during transcription

# Transcribe with speaker identification
crispasr --backend parakeet \
         --speaker-db ./speakers \
         -f meeting.wav
# Output: (alice) Hello everyone...

Standalone embedding extraction

# Extract speaker embedding (test binary)
test-titanet titanet-large.gguf audio1.wav audio2.wav
# Prints cosine similarity matrix

Python

from crispasr import TitaNet, SpeakerDB

with TitaNet("titanet-large.gguf") as model:
    emb = model.embed(pcm_16k_float32)

db = SpeakerDB("./speakers")
db.enroll("alice", emb)
name, score = db.match(emb, threshold=0.7)

Conversion

python models/convert-titanet-to-gguf.py \
    --input nvidia/speakerverification_en_titanet_large \
    --output titanet-large.gguf

Verification

Encoder + decoder parity with NeMo reference: cos = 0.999997 (mel-injected). End-to-end parity: cos = 0.917 (STFT float32 precision gap in mel front-end).

Citation

@inproceedings{koluguri2022titanet,
  title={TitaNet: Neural Model for Speaker Representation with 1D Depth-wise Separable Convolutions and Global Context},
  author={Koluguri, Nithin Rao and Park, Taejin and Ginsburg, Boris},
  booktitle={ICASSP 2022},
  year={2022}
}