adolforeyna/whisper-translation
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Switch to dedicated MarianMT for cleaner, non-LLM translation

Browse Source
This commit is contained in:
Adolfo Reyna
2026-02-26 21:13:03 -05:00
parent 6f602a1868
commit 8e45daec87
1 changed files with 17 additions and 17 deletions
Download Patch File Download Diff File Expand all files Collapse all files

34
transcribe.py
View File

@@ -5,12 +5,12 @@ import queue
import sys
import torch
from silero_vad import load_silero_vad, get_speech_timestamps
from mlx_lm import load, generate
from transformers import MarianMTModel, MarianTokenizer
# Parameters
WHISPER_MODEL = "mlx-community/whisper-small.en-mlx"
LLM_MODEL = "mlx-community/SmolLM2-135M-Instruct" # Verified public model
TARGET_LANG = "Spanish"
# Dedicated EN-ES translation model (very fast and accurate)
TRANSLATE_MODEL = "Helsinki-NLP/opus-mt-en-es"
CHANNELS = 1
SAMPLERATE = 16000
BLOCK_SIZE = 512
@@ -25,26 +25,23 @@ def callback(indata, frames, time, status):
print(status, file=sys.stderr)
audio_queue.put(indata.copy())
def translate_text(model, tokenizer, text, target_lang):
prompt = f"Translate the following English text to {target_lang}. Only provide the translation, no extra text.\n\nEnglish: {text}\n\n{target_lang}:"
messages = [{"role": "user", "content": prompt}]
prompt_text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
response = generate(model, tokenizer, prompt=prompt_text, max_tokens=100, verbose=False)
return response.strip()
def main():
# Set device for translation model
device = "mps" if torch.backends.mps.is_available() else "cpu"
print(f"Using device for translation: {device}")
print(f"Loading Whisper model '{WHISPER_MODEL}'...")
print(f"Loading LLM for translation '{LLM_MODEL}'...")
llm_model, llm_tokenizer = load(LLM_MODEL)
print(f"Loading dedicated translation model '{TRANSLATE_MODEL}'...")
tokenizer = MarianTokenizer.from_pretrained(TRANSLATE_MODEL)
model = MarianMTModel.from_pretrained(TRANSLATE_MODEL).to(device)
print("Loading Silero VAD model...")
vad_model = load_silero_vad()
print("Models loaded.")
print(f"\nStarting live transcription & translation to {TARGET_LANG}... (Press Ctrl+C to stop)")
print(f"\nStarting live transcription & translation... (Press Ctrl+C to stop)")
audio_buffer = []
speech_started = False
@@ -80,11 +77,14 @@ def main():
original_text = result['text'].strip()
if original_text:
# 2. Translate using LLM
translated_text = translate_text(llm_model, llm_tokenizer, original_text, TARGET_LANG)
# 2. Translate using dedicated MarianMT
inputs = tokenizer(original_text, return_tensors="pt", padding=True).to(device)
with torch.no_grad():
translated_tokens = model.generate(**inputs)
translated_text = tokenizer.decode(translated_tokens[0], skip_special_tokens=True)
print(f"\nEN: {original_text}")
print(f"{TARGET_LANG[:2].upper()}: {translated_text}")
print(f"ES: {translated_text}")
audio_buffer = []
speech_started = False