Nimruz/Nava-ct2

Nava for faster-whisper

Nava, a Persian speech-to-text model, converted to CTranslate2 (float16) for faster-whisper. Results, training details and limitations are on the main model page.

from faster_whisper import WhisperModel

model = WhisperModel("Nimruz/Nava-ct2", device="cuda", compute_type="float16")
segments, _ = model.transcribe(
    "audio.wav", language="fa", beam_size=1,
    without_timestamps=True, condition_on_previous_text=False,
)
print(" ".join(s.text.strip() for s in segments))

On CPU use device="cpu", compute_type="float32", or "int8" for less memory.

Things to know

• Always pass language="fa" and without_timestamps=True. Nava was fine-tuned with the Persian prompt and without timestamp tokens.
• Audio longer than 30 seconds needs splitting. Cut it at quiet points into pieces of at most 30 seconds, or use the transcribe.py script in the main repo. vad_filter=True should serve the same purpose but has not been tested with Nava.
• Output has no punctuation and spells numbers out. The main repo has a formatter that turns spoken numbers, dates and times into digits.

This conversion gave the same text as the original model on 9 of 10 FLEURS test clips, with a difference of a word or two on the tenth.