Text to speech

​

OpenAI SDK

from openai import OpenAI
client = OpenAI(base_url="https://tts.eu.mihu.ai/v1", api_key=KEY)

client.audio.speech.create(model="tts-1", voice="alloy", input="Merhaba, siparişiniz yolda.").write_to_file("out.mp3")

with client.audio.speech.with_streaming_response.create(
        model="gpt-4o-mini-tts", voice="onyx", input="Guten Tag! Ihre Bestellung ist unterwegs.",
        response_format="opus", speed=1.1) as r:
    for chunk in r.iter_bytes():
        ...   # play or forward while it streams

​

Streaming

• Audio is sent as soon as it is generated. With pcm the first bytes are audio; with wav a 44-byte header comes first.
• mp3, opus, aac, flac and speed other than 1 go through an encoder that adds about 0.07 s to the first audio.
• If the client closes the connection (for example on barge-in), synthesis stops.
• With stream_format: "sse" the answer is Server-Sent Events: speech.audio.delta events with base64 audio, then speech.audio.done.

​

Answer headers

​

Text handling

• Normalization (normalize, on by default): numbers, amounts, dates, times and abbreviations are written out (“14.30” → “on dört otuz”). false reads the text as it is.
• Generate and verify (verify, Turkish, off by default): a speech recognizer listens to each sentence chunk; a chunk it does not hear as written is said once more and the better take is sent. Fewer misread or skipped words, but the first audio comes about 2 s later. For files and prepared announcements, not for live calls.
• Russian: stress marks are added automatically. Text that already carries them (U+0301 after the vowel, or + before it: з+амок) is kept as written. Latin words are said in Cyrillic; for a name that must sound a certain way, write it in Cyrillic.

​

Errors