index · glossary
11 pages
Voice-agent terms, defined.
Definitions written the way we use them in benchmarks — each term with the honest way to measure it, and the ways vendors blur it.
01
107 / 108
ms, p50 / p95 — one millisecond of spread
Time to first audio (TTFA)
TTFA is the delay from sending text to the first playable audio byte. How it differs from model latency and TTFB, and how to measure it honestly.
02
~100 ms
first chunk, playback overlapping synthesis
Streaming text-to-speech
Streaming TTS returns audio in chunks as synthesis proceeds instead of one file at the end. Why voice agents require it, and what to check in a streaming API.
03
10 s
of reference — the whole enrollment
Instant voice cloning
Instant cloning builds a voice from a ten-second reference at request time — no training job. What zero-shot means, and how identity survives translation.
04
×2
the latency budget, paid twice per interruption
Barge-in
Barge-in is the caller interrupting the agent mid-sentence. What the TTS layer must do to support it: stop instantly, resynthesize fast, stay cheap.
05
1 line
one conversation, right now
Concurrent lines
A line is one simultaneous conversation. Why concurrency — not minutes — is the real capacity unit for voice agents, and how metered plans cap it quietly.
06
< 1
the bar any streaming voice must clear
Real-time factor (RTF)
RTF is synthesis time divided by audio duration. Why RTF below 1 is table stakes, and why time to first audio matters more for conversations.
07
1 shape
the same request on every transport
SSE vs WebSocket for TTS
When to stream synthesis over server-sent events and when to hold a WebSocket: connection lifetime, turn structure, and infrastructure fit.
08
$30–$100
published meters, per million characters
TTS pricing: per character vs flat rate
The two ways speech synthesis is priced, what each one does to a voice product, and which side of the break-even your workload sits on.
09
~1.5 MB
the reference cap — about 30 seconds
Voice cloning reference audio
The reference is the clip a zero-shot clone is built from: about ten seconds of one clean, consenting speaker, a ~1.5 MB cap, fingerprinted and cached.
10
loaded
idling in memory until the fleet fills
Warm spill
Warm spill is a second serving tier that is already running when the primary fleet fills, so overflow traffic costs latency instead of availability.
11
100%
of samples marked at generation
Audio watermarking
Audio watermarking marks synthesized speech as synthetic. How an inaudible watermark survives where metadata does not, and what a watermark cannot do.