SOV-Transcribe-1

Salesforce's 26.6B parameter speech model combines Whisper large-v3 with a Gemma 4 MoE decoder (4.6B active parameters). It transcribes English, French, German, Spanish, Italian, Portuguese, and Dutch, with natural-language instructions and contextual vocabulary. Recordings can be up to 30 minutes long.

MLX checkpoint	Download size
salesforce-sovereign-intelligence/SOV-Transcribe-1 (official BF16)	53.3 GB
nativ-community/SOV-Transcribe-1-8bit	28.2 GB
nativ-community/SOV-Transcribe-1-4bit	15.6 GB

Both checkpoints preserve the audio encoder and projector in BF16 and the MoE routers in FP32.

from mlx_audio.stt import load

model = load("nativ-community/SOV-Transcribe-1-4bit")
result = model.generate(
    "audio.wav",
    prompt="Transcribe in written form with punctuation.",
    hotwords=["Salesforce", "MLX"],
)
print(result.text)

For incremental text after audio encoding, pass stream=True:

for chunk in model.generate("audio.wav", stream=True):
    print(chunk.text, end="", flush=True)

audio_batch_size controls how many 30-second windows are encoded together (default 1); prefill_step_size controls decoder prefill (default 512).

mlx_audio.stt.generate \
  --model nativ-community/SOV-Transcribe-1-4bit \
  --audio audio.wav --output-path transcript --verbose

The original BF16 checkpoint also loads directly and occupies about 53 GB on disk. Audio files are resampled to 16 kHz; waveform arrays must already be mono at 16 kHz. Transcription runs offline. The returned finish_reason indicates whether decoding stopped normally or reached max_tokens.

See the model README for additional usage and license information, and the upstream model for model details.