nativ-community/SOV-Transcribe-1-8bit

SOV-Transcribe-1 (8-bit)

MLX conversion of Salesforce SOV-Transcribe-1, combining a Whisper large-v3 audio encoder with a Gemma 4 mixture-of-experts decoder: 26.6B total parameters and 4.6B active parameters.

The model supports offline transcription with natural-language instructions, language hints, and contextual vocabulary in English, French, German, Spanish, Italian, Portuguese, and Dutch. Input recordings can be up to 30 minutes long.

Quantization

Component	Precision
Decoder embedding, attention, dense MLPs, and experts	8-bit affine, group size 64
Audio encoder and audio projector	BF16
Mixture-of-experts routers	FP32

Usage

Requires a version of mlx-audio with SOV-Transcribe-1 support.

from mlx_audio.stt import load

model = load("nativ-community/SOV-Transcribe-1-8bit")
result = model.generate("audio.wav")
print(result.text)

result = model.generate(
    "meeting.wav",
    prompt="Transcribe in written form with punctuation and capitalization.",
    hotwords=["Salesforce", "MLX"],
    max_tokens=8192,
)
print(result.text)

mlx_audio.stt.generate \
  --model nativ-community/SOV-Transcribe-1-8bit \
  --audio audio.wav \
  --output-path transcript \
  --verbose

Audio files are resampled to 16 kHz. Waveform arrays must already be mono at 16 kHz. Results include token counts, timing statistics, and finish_reason (stop or length). Increase max_tokens if decoding reaches its limit.

Validation

Loaded locally with strict weight checking and tested on seven smoke cases: two English voices, French, a separate English passage, a 66.92-second recording with speech before and after a 47-second silence using both chunked and single prefill, and a transcription request with instructions and hotwords. Every reference word matched after punctuation normalization; every case stopped normally. The two prefill strategies produced identical long-form transcripts. This small smoke set is not a general recognition benchmark.

Source and license

Converted from revision 505d8fae243924c0d8acdc622f7f0abab344eba3 of salesforce-sovereign-intelligence/SOV-Transcribe-1. The model and Salesforce reference code use the Apache 2.0 license.