Put a pause marker in your text and Flux TTS waits. Spell a word in IPA and it says it your way (Early Access).
Since September 30, 2026, Deepgram's Flux text-to-speech (/v2/speak) reads two kinds of marker placed inside the text you send: one that inserts a silence, and one that tells it how to pronounce a word.
Wirestack
DEEPGRAM_API_KEY (paste it where the examples say YOUR_API_KEY)pip install deepgram-sdk)\{pause:800ms\} where the voice should wait. Pauses run from 500 ms to 3000 ms in 100 ms steps, up to 8 per request, on batch (REST) requests:from deepgram import DeepgramClient
client = DeepgramClient(api_key="YOUR_API_KEY")
text = r"Your appointment is on Tuesday at 3 PM. \{pause:800ms\} Reply YES to confirm."
response = client.speak.v2.audio.generate(
text=text,
model="flux-haley-en",
encoding="mp3",
speed=0.9,
)
audio_bytes = b"".join(response)
with open("appointment_reminder.mp3", "wb") as f:
f.write(audio_bytes)from deepgram import DeepgramClient
client = DeepgramClient(api_key="YOUR_API_KEY")
text = r'Take \{"word": "Azathioprine", "pronounce": "æzəˈθaɪəpɹiːn"\} twice daily with \{"word": "dupilumab", "pronounce": "duːˈpɪljuːmæb"\}.'
response = client.speak.v2.audio.generate(
text=text,
model="flux-haley-en",
encoding="mp3",
)
audio_bytes = b"".join(response)
with open("medical_instructions.mp3", "wb") as f:
f.write(audio_bytes)speed other than 1.0. With a pause present, speed is capped at 1.15. Combinations outside these rules are rejected with CONTROL_COMBINATION_INVALID or PAUSE_SPEED_CAP_EXCEEDED.Run the first script and play appointment_reminder.mp3: there should be a clear gap before "Reply YES to confirm." Deepgram notes that pronunciation results can vary while the feature is in Early Access, so generate each term a few times before relying on it.