OpenAI's smallest Whisper model: transcribes and translates speech in about 100 languages. 39M parameters.
About
Whisper is an encoder–decoder speech model trained by OpenAI on 680,000 hours of labelled audio. Tiny is the smallest size: quick enough for a CPU, at the cost of noticeably more errors than the larger versions. It transcribes speech in its own language or translates it into English, and works on 30-second windows of 16 kHz audio.
Use it
from transformers import pipeline
transcribe = pipeline("automatic-speech-recognition", model="openai/whisper-tiny")
transcribe("meeting.wav")
Limitations
Accuracy varies a lot by language, accent and recording quality, and it can "hallucinate" text that was never said, especially over silence or music. Don't use it where transcription errors have serious consequences.
Licence
Apache-2.0. Files link to OpenAI's repository on Hugging Face.