99 languages, 48 voices, 128ms latency — how to pick the right speech model for voice assistants, transcription, and multilingual apps.
Speech workloads fail on different axes than text ones. Latency budgets are hard, language coverage is uneven, and the failure mode users notice most is not a wrong word but an awkward pause.
Start with the latency budget
A voice assistant needs first audio out in under 300ms end to end. That leaves roughly 120ms for TTS once you account for ASR, the model turn, and the network. Batch transcription has no such budget and should optimize purely for accuracy per dollar.
- Interactive voice: streaming ASR + 128ms TTS, accept a higher word error rate
- Meeting transcription: batch ASR with diarization, accuracy first
- Multilingual IVR: coverage first — verify your exact locale, not just the language
- Audiobook and narration: highest-quality TTS, latency irrelevant, pick on voice range
Language coverage is not uniform
99 supported languages does not mean 99 equally good languages. Word error rate on our top ten locales averages 3.1%; across the long tail it is 11.4%.
Always benchmark on your own audio. Twenty minutes of representative recordings — real microphones, real background noise, real accents — will tell you more than any published benchmark.
Streaming both directions
with client.audio.speech.with_streaming_response.create(
model="gpt-4o-mini-tts",
voice="aria",
input=text,
extra_body={"amplink": {"first_chunk_target_ms": 128}},
) as stream:
for chunk in stream.iter_bytes():
play(chunk)