mlx-community/diar_streaming_sortformer_4spk-v2.1-fp32

This model was converted to MLX format from nvidia/diar_streaming_sortformer_4spk-v2.1 using mlx-audio version 0.3.2.

Refer to the original model card for more details on the model.

Use with mlx-audio

pip install -U mlx-audio

Converting from NeMo

The original model is distributed as a .nemo archive. This repo contains the pre-converted MLX weights.

python -m mlx_audio.vad.models.sortformer.convert \
    --nemo-path nvidia/diar_streaming_sortformer_4spk-v2.1 \
    --output-dir ./sortformer-v2.1-mlx

Python Example — Streaming Inference (Recommended):

from mlx_audio.vad import load

model = load("mlx-community/diar_streaming_sortformer_4spk-v2.1-fp32")

for result in model.generate_stream("meeting.wav", chunk_duration=5.0, verbose=True):
    for seg in result.segments:
        print(f"Speaker {seg.speaker}: {seg.start:.2f}s - {seg.end:.2f}s")

Python Example — Offline Inference:

from mlx_audio.vad import load

model = load("mlx-community/diar_streaming_sortformer_4spk-v2.1-fp32")
result = model.generate("meeting.wav", threshold=0.5, verbose=True)
print(result.text)

Python Example — Real-time Microphone Streaming:

from mlx_audio.vad import load

model = load("mlx-community/diar_streaming_sortformer_4spk-v2.1-fp32")
state = model.init_streaming_state()

for chunk in mic_stream():  # your audio source
    result, state = model.feed(chunk, state, sample_rate=16000)
    for seg in result.segments:
        print(f"Speaker {seg.speaker}: {seg.start:.2f}s - {seg.end:.2f}s")

Model Details

• Architecture: FastConformer (17 layers) + Transformer Encoder (18 layers) + Sortformer Modules
• Mel bins: 128
• Max speakers: 4
• Streaming: AOSC (Arrival-Order Speaker Cache) compression for intelligent long-range context
• Input: 16kHz mono audio
• Output: Per-frame speaker activity probabilities

Key Streaming Features

• Speaker Cache + FIFO buffers for long-range and recent context
• AOSC compression scores frames by per-speaker log-likelihood ratio, boosting underrepresented speakers
• Silence profiling fills cache gaps with running-mean silence embeddings
• Left/right context for chunk boundary handling in file mode