Build Speech-to-Text with MAI-Transcribe-1.5

# %pip install -q requests python-dotenv

import os
import json
from copy import deepcopy
from pathlib import Path
import requests
from dotenv import load_dotenv

ENV_PATH = 'deployment.env' if os.path.exists('deployment.env') else os.path.join('..', 'deployment.env')
load_dotenv(ENV_PATH, override=True)

SPEECH_ENDPOINT = os.getenv('MAI_TRANSCRIBE_15_ENDPOINT', 'https://eastus.api.cognitive.microsoft.com/').rstrip('/')
SPEECH_KEY = os.getenv('MAI_TRANSCRIBE_15_KEY')
TRANSCRIBE_URL = f"{SPEECH_ENDPOINT}/speechtotext/transcriptions:transcribe?api-version=2025-10-15"
assert SPEECH_KEY, 'Set MAI_TRANSCRIBE_15_KEY in deployment.env'

audio_dir_env = os.getenv('TRANSCRIBE_LOCAL_AUDIO_DIR')
LOCAL_AUDIO_DIR = Path(audio_dir_env) if audio_dir_env else Path('media') / 'mai-transcribe-1-5'
assert LOCAL_AUDIO_DIR.exists(), f'Audio folder not found: {LOCAL_AUDIO_DIR}'
candidates = sorted(LOCAL_AUDIO_DIR.glob('*.wav')) + sorted(LOCAL_AUDIO_DIR.glob('*.mp3')) + sorted(LOCAL_AUDIO_DIR.glob('*.flac'))
non_empty = [p for p in candidates if p.stat().st_size > 0]
assert non_empty, f'No non-empty WAV/MP3/FLAC files found in: {LOCAL_AUDIO_DIR}'
AUDIO_FILE = str(max(non_empty, key=lambda p: p.stat().st_size))
print('Endpoint:', SPEECH_ENDPOINT)
print('Audio file:', AUDIO_FILE)

def _mime_for(path: Path) -> str:
    s = path.suffix.lower()
    if s == '.wav':
        return 'audio/wav'
    if s == '.mp3':
        return 'audio/mpeg'
    if s == '.flac':
        return 'audio/flac'
    return 'application/octet-stream'

def transcribe_with_definition(audio_path: str, definition: dict) -> dict:
    p = Path(audio_path)
    with p.open('rb') as f:
        files = {
            'audio': (p.name, f, _mime_for(p)),
            'definition': (None, json.dumps(definition), 'application/json'),
        }
        resp = requests.post(
            TRANSCRIBE_URL,
            headers={'Ocp-Apim-Subscription-Key': SPEECH_KEY},
            files=files,
            timeout=300,
        )
    if resp.ok:
        return resp.json()

    # Compatibility fallback for backends that reject enhancedMode/model
    txt = (resp.text or '').lower()
    if resp.status_code == 400 and 'enhanced mode with model' in txt:
        fb = deepcopy(definition)
        fb.get('enhancedMode', {}).pop('model', None)
        return transcribe_with_definition(audio_path, fb)
    if resp.status_code == 400 and 'enhanced mode is currently not supported' in txt:
        fb = deepcopy(definition)
        fb.pop('enhancedMode', None)
        return transcribe_with_definition(audio_path, fb)

    raise requests.HTTPError(f'Transcription failed with {resp.status_code}: {resp.text}', response=resp)

general_definition = {
    'enhancedMode': {
        'enabled': True,
        'model': 'mai-transcribe-1.5'
    }
}
general_result = transcribe_with_definition(AUDIO_FILE, general_definition)
print(json.dumps(general_result, indent=2)[:2000])

verbatim_definition = {
    'enhancedMode': {
        'enabled': True,
        'model': 'mai-transcribe-1.5',
        'transcribeStyle': 'verbatim'
    }
}
verbatim_result = transcribe_with_definition(AUDIO_FILE, verbatim_definition)
print(json.dumps(verbatim_result, indent=2)[:2000])

phrase_list_definition = {
    'phraseList': {
        'phrases': ['MAI', 'Microsoft Build', 'Azure Speech', 'Foundry', 'Copilot']
    },
    'enhancedMode': {
        'enabled': True,
        'model': 'mai-transcribe-1.5'
    }
}
phrase_list_result = transcribe_with_definition(AUDIO_FILE, phrase_list_definition)
print(json.dumps(phrase_list_result, indent=2)[:2000])

lid_definition = {
    'locales': ['en-US', 'es-ES', 'fr-FR', 'de-DE'],
    'enhancedMode': {
        'enabled': True,
        'model': 'mai-transcribe-1.5'
    },
}
lid_result = transcribe_with_definition(AUDIO_FILE, lid_definition)
print(json.dumps(lid_result, indent=2)[:2000])