Best TTS for real-time voice agents

TTFB leaderboard

Time-to-first-byte on each vendor's lowest-latency streaming model. US-East origin, 40-char prompt, measured over 50 calls in April 2026. Piper runs locally on an M2 laptop CPU.

The end-to-end round trip

TTS is one of five hops between user voice in and synthesized voice out. Optimizing TTS alone without a fast STT and a streaming LLM is pointless.

Quality doesn't have to drop

Fast doesn't mean bad. The upper-left of this plot (fast + natural) is populated: Sonic 2 and Flash v2.5 both clear 4.5 MOS at sub-100ms. Piper is the only budget option; the rest cluster in the 4.2-4.7 range.

Voice bot reference stack (2026)

Streaming setup (Cartesia)

# Cartesia Sonic 2 over WebSocket — the lowest-latency setup in production today.
from cartesia import Cartesia
client = Cartesia(api_key="sk_...")
ws = client.tts.websocket()

# Incremental text -> incremental PCM. Send tokens as your LLM produces them.
def stream_tokens(llm_stream):
    for token in llm_stream:
        ws.send(
            model_id="sonic-2",
            transcript=token,
            voice={"mode": "id", "id": "<voice_id>"},
            output_format={"container": "raw", "encoding": "pcm_s16le", "sample_rate": 24000},
            continue_=True,
        )
    ws.send(transcript="", continue_=False)  # flush

    for chunk in ws.receive():
        speaker.write(chunk.audio)

Note the continue_=True flag. You want to send LLM tokens as they arrive rather than waiting for a full sentence — this collapses the LLM→TTS serial delay into a single pipeline.

Listen: same prompt, every vendor