Levandor

STT Vendor Landscape & the Telephony Benchmark Gap

5 min read

STT selection for narrowband (8kHz μ-law) inbound telephony. Deepgram Flux is the pick — but the more important finding is that every published streaming-STT leaderboard is measured on audio a phone trunk will never deliver.

Deepgram Flux — The Pick

flux-general-en, endpoint wss://api.deepgram.com/v2/listen.

Note the /v2 — Flux is not on /v1

This trips up SDK integration. Everything else in the Deepgram surface (Nova-2, Nova-3) is /v1/listen. Flux is a v2 endpoint.

Why it wins

  1. Fused end-of-turn detection. Flux folds EOT detection into the acoustic model itself rather than bolting a separate VAD on top. ~260ms EOT p50.
  2. Native encoding=mulaw&sample_rate=8000. Twilio Media Streams frames go in with zero transcoding — no resample, no format conversion, no added buffer. This is a latency and a code-complexity win.
  3. EagerEndOfTurn / TurnResumed event pair. Fires 150–250ms before confirmed EOT.
  4. The only first-party Rust SDK in the entire field. Crate deepgram v0.10.0, published 2026-07-12. Deepgram also publishes a Twilio↔Rust streaming tutorial. For a Rust-native build where every other vendor is Python/Node-first, this is a real advantage.

Flux tail is long — p50 is not the story

PercentileEOT latency
p50~260ms
p90~1,000ms
p95~1,500ms

This is consistent with the general finding that every platform roughly triples p50→p95. Budget against p95.

Speculative LLM dispatch

Use EagerEndOfTurn to hide LLM TTFT

Dispatch the LLM on EagerEndOfTurn; cancel on TurnResumed. This buys back 150–250ms of the ~270ms LLM TTFT.

Cost: 50–70% additional LLM calls. At Haiku 4.5 prices with prompt caching (~$0.0032/min, ~10% of total cost) this is trivially worth it — a 70% increase on 10% of cost is a ~7% cost increase for a ~200ms latency win. See voice-agent-cost-model-2026-07-21.

Tuning knobs

ParameterRangeDefault
eot_threshold0.5 – 0.90.7
eager_eot_threshold0.3 – 0.9—
eot_timeout_ms500 – 60,000—

Lower eot_threshold = faster commit, more mid-sentence cut-offs. For a receptionist taking phone numbers and addresses this needs to be context-dependent — the digit-string case is exactly the [[voice-agent-managed-platform-control-surfaces-2026-07-21#the-onnumberseconds-gap|onNumberSeconds problem]] that only two managed platforms solve, and which owning the pipeline lets you solve properly.

Verify before building

[unverified] whether the deepgram Rust crate v0.10.0 covers /v2/listen (Flux) or only /v1. The crate is the headline reason to prefer Deepgram in a Rust build; if it’s v1-only, that advantage partly evaporates and you’re hand-rolling the WebSocket client anyway.

AssemblyAI Universal-Streaming — Cheaper, But Slow to Commit

PropertyValue
Price$0.0025/min — cheaper than Flux
Word emission P50307ms
max_turn_silence default1,280ms

~5x slower than Flux to commit a turn

Word emission latency (307ms) looks competitive, but that’s words appearing, not the turn being committed. The 1,280ms default max_turn_silence means the agent waits over a second of silence before deciding the caller is done. That is past the ~800ms “unnatural” threshold on its own, before the LLM has even started.

Billing gotcha — bills by SESSION duration, not audio duration

AssemblyAI bills the wall-clock length of the WebSocket session, not the amount of audio you send. For telephony, where the connection stays open across long silences and hold periods, this makes the headline $0.0025/min misleading. Model it against real call duration, not real speech duration.

Critical Benchmark Caveat

Artificial Analysis' streaming STT index contains NO narrowband telephony audio at all

Composition of the index:

  • 50% AA-AgentTalk
  • 25% VoxPopuli (European Parliament recordings)
  • 25% Earnings22 (earnings calls)

All wideband. Every published leaderboard position is measured on audio a G.711 trunk will never deliver. Narrowband 8kHz μ-law drops everything above ~3.4kHz and adds companding noise — the exact band where fricative and sibilant discrimination lives (s/f, p/t/k). This is where telephony WER diverges from benchmark WER, and it’s the band that matters most for names, emails, and spelled-out strings.

Nobody publishes per-domain telephony WER

This is a genuine gap in the public record, not an oversight in the research. There is no leaderboard that answers “which STT is best on 8kHz μ-law.”

The required mitigation

Budget a week for an own A/B on real trunk audio

Record 8kHz μ-law directly off the actual Twilio trunk — not resampled studio audio, not a synthetic downsample. Then A/B candidate models on it.

Use Deepgram Nova-2 phonecall as the control. It is the only model in the field explicitly special-cased for narrowband. If Flux can’t beat Nova-2 phonecall on real trunk audio, the fused-EOT advantage isn’t paying for the accuracy loss.

Weight the test set toward the failure modes that matter for a receptionist: names, email addresses, phone numbers, booking references, spelled-out alphanumerics. These are the same categories that kill speech-to-speech models and they will be the weakest part of any cascade too.

This is open item #6 in voice-agent and the highest-value one — it’s the only item that could change the vendor decision.

Decision Summary

Deepgram FluxAssemblyAI Universal-StreamingDeepgram Nova-2 phonecall
Turn commit~260ms p50~1,280ms defaultn/a (external VAD)
Native μ-law 8kyes—yes, tuned for it
Rust SDKfirst-party—first-party
Speculative dispatchEagerEndOfTurn——
Price$0.0065/min$0.0025/min (session-billed)—
Roleprimaryrejected on commit latencyA/B control

Sources

  • Deepgram Flux docs — endpoint, encoding, EagerEndOfTurn, tuning params [vendor]
  • Deepgram Rust crate v0.10.0, published 2026-07-12 [vendor]
  • Deepgram Twilio↔Rust streaming tutorial [vendor]
  • AssemblyAI Universal-Streaming docs — max_turn_silence, session billing [vendor]
  • Artificial Analysis streaming STT index — dataset composition [measured, but wideband-only]

Graph View

Backlinks