docs/LATENCY.md

Latency methodology

How the latency numbers in the card and LEADERBOARD.md were measured.

What we measure

A voice agent waits after the customer stops speaking: first for its voice-activity detector (VAD) to decide the turn has ended, then for the speech-to-text system to return the turn's final text. The VAD wait is the same whichever STT system is used; the second wait is what the STT system adds. We measure that second wait directly.

FTR — Finalize-To-Response (headline). Every call is streamed to the system in real time. At each end of a customer turn, as decided by our production VAD, we send the system its finalize signal (its standard way of being told the turn is over) and record the time. FTR is the time from sending finalize to receiving the system's reply for that turn. It is measured for every turn and pooled across the set.

TTFS — Time To Final Segment. The time from when the customer actually stopped speaking to the final transcript, as defined by pipecat-ai/stt-benchmark. Our finalize is sent at the VAD stop, and the VAD stop fires a fixed 0.48 s after speech ends, so here TTFS = FTR + 480 ms (measured gap: median 480.5 ms).

Our VAD

The same voice-activity detector as our production voice agent, with a stop of 0.5 s (an effective stop of 0.48 s at 8 kHz). Turn boundaries are computed once per call and are identical for every system.

How this differs from Pipecat's TTFS

Pipecat measures from speech end, so its TTFS includes the VAD stop wait. Its benchmark defaults to vad_stop_secs = 0.2 (the CLI default, which its docs recommend keeping; its results workflow runs with the default), so its published TTFS most likely includes a 0.2 s stop; the value is not recorded per row. FTR excludes the VAD wait and isolates what the STT system adds. Because the VAD settings differ (0.48 s here, 0.2 s by default there), our TTFS and Pipecat's are not directly comparable; FTR is comparable across VAD settings. Pipecat also times one utterance per sample; we time every turn of full, unsegmented calls.

Which reply counts

Reply type	Systems	How FTR is timed
Tagged	Nova 3 (from_finalize), Arth V1/V2 (from_finalize), Soniox v5 (<fin> token), Scribe v2 (k-th committed transcript), Cartesia ink-2 (flush_done), Flux (EndOfTurn, trigger=manual), OpenAI (item_id of the committed audio)	The system labels the message that answers our finalize, so FTR is exact. Turns with no speech get an empty reply, which still counts: it is the system's "turn complete" signal.
Inferred	Saaras v4, Gemini 3.5 Transcribe	No label, so the reply is the first final received after finalize and before the next turn starts. Comparable on turns with speech (counting only replies that carry text moves FTR by 1–5 ms for every system), but these systems give no explicit "turn complete" signal.
Not applicable	Chirp 3	No client-side finalize (only its own endpointing), so FTR does not apply.

Setup

• Calls: a locked set of 50 calls from the benchmark (70 min of audio, 34–202 s each; 541 VAD turns), stratified by duration (10 buckets × 5 calls, fixed seed), each transcribed by every system in the full run.
• Streaming: 8 kHz mono audio sent in real time, one stream per call.
• Load: 5 concurrent streams, one system at a time, all from the same server in Delhi, India. Client-side pacing lag was recorded per call and was small.
• Language: Hindi, or a Hindi + English hint where the API takes a list; never auto-detect. Strict-language options were used where they exist.
• Configuration notes:
  • Nova 3 is measured on Deepgram's India endpoint (api.in.deepgram.com) with server endpointing off (endpointing=false), so that, like the others, it returns its text in reply to our finalize. With Deepgram's endpointing on, the text usually arrives before the finalize and FTR would only time an empty acknowledgement.
  • Soniox v5 is measured on its India regional endpoint (see below).
  • Cartesia ink-2 is measured on its India regional endpoint (api-india.cartesia.ai).

Results

Locked 50-call set; milliseconds; lower is better.

System	Region	Hosting	FTR P50	P80	P90	P95	P99	TTFS P50
Soniox v5	India	Provider cloud	55.6	89.4	97.2	100.2	108.0	536
Arth V1 ★	India	Self-hosted	62.5	67.0	70.9	74.3	84.8	543
Arth V2 ★	India	Self-hosted	62.9	67.0	70.9	75.7	88.7	543
Scribe v2 Realtime	India	Provider cloud	64.8	86.5	111.6	134.3	248.1	545
Nova 3	India	Provider cloud	73.7	93.8	104.2	114.2	140.0	554
Flux	India	Provider cloud	113.8	124.4	128.7	132.6	178.5	594
Cartesia ink-2	India	Provider cloud	144.6	183.6	202.3	218.5	260.7	625
Saaras v4	India	Provider cloud	149.6	176.7	204.3	228.3	671.6	630
Gemini 3.5 Transcribe	not documented	Provider cloud	243.2	264.2	281.0	304.3	335.9	723
OpenAI gpt-live-transcribe	not documented	Provider cloud	766.5	830.8	876.9	919.0	1235.6	1247

Self-hosted: SquadStack's own Arth deployments in India. Chirp 3 is not listed (no client-side finalize).

Region matters: Soniox v5 from one server

The same Soniox model and configuration, measured from the same Delhi server against each regional endpoint:

Soniox v5 endpoint	FTR P50	P95	WebSocket round trip from the server
India (stt-rt.in.soniox.com)	55.6	100.2	23 ms
Japan (stt-rt.jp.soniox.com)	204.6	254.0	181 ms
EU (stt-rt.eu.soniox.com)	228.8	290.0	234 ms
US (stt-rt.soniox.com)	292.5	345.5	272 ms

FTR is about one network round trip plus 15–40 ms of processing in every region, so for hosted APIs the region largely decides the number. A fresh re-run on the India endpoint reproduced it (P50 56.0, P95 101.3 ms), and transcripts across the four regions differ by 1.6–2.5% of words, within the run-to-run variation of a single region.

Caveats

• All numbers are from one server location in India. Systems served from India have a network advantage; reproduce from your own location before relying on absolute values.
• OpenAI streams the full turn text as incremental deltas about 400 ms before the final completed event that FTR measures.
• Gemini runs in manual activity mode; the next turn's start is held for up to 1 s until the previous reply arrives, so that replies are not merged across turns.
• 50 calls give stable P50/P95; P99 rests on a handful of turns.