Neural TTS Latency Benchmarks Across Leading APIs

• Role: Opens the conceptual foundation the rest of the piece builds on — establishes TTFA/TTFB as the production-relevant metric before any benchmark numbers appear, so readers can evaluate every figure that follows with the right frame.
• TTFA (Time to First Audio) / TTFB (Time to First Byte): elapsed time from API receiving text to delivering first audio chunk to the client
• Does not measure time to complete audio — that scales with text length and is irrelevant to when playback can begin
• Why this matters perceptually: human turn-taking gaps run 100–300ms; silence beyond that begins to feel awkward, with a truly broken experience typically perceived above ~1 second, not a slow one
• Contrast with other metrics developers may reach for first:
  • Total generation time — useful for batch jobs, misleading for agents
  • Real-Time Factor (RTF) — describes synthesis speed relative to playback speed, relevant for headroom; an RTF of 0.3 means audio generates 3.3× faster than it plays
  • Word Error Rate (WER) — accuracy dimension, orthogonal to latency
• Streaming vs. batch as an architectural fork: streaming (WebSocket or SSE) returns audio chunks as they generate; batch builds the whole file first — the streaming/batch choice determines whether TTFA is even the right question
• Neural vocoders now dominate: neural voice technology holds more than 83% of TTS market share per Deepgram's 2026 guide, making this the baseline architecture for every provider compared later
• No benchmark number yet — this section only builds the interpretive lens

Why the same API can show radically different numbers depending on how you measure it

• Role: Complicates the apparently simple "here are the numbers" premise — readers must understand measurement conditions before they can use any figure, including the ones about to appear in the provider section.
• Vendor-reported figures are typically measured under controlled, low-contention conditions that do not resemble production traffic
• What vendor numbers omit:
  • Network path and geographic distance from client to inference node
  • Connection setup overhead — WebSocket vs. HTTP per-request adds 40–100ms per turn in multi-turn conversations
  • Text normalization time before phoneme prediction begins
  • Client-side buffering and playback start delay
  • Upstream pipeline: in a voice agent, STT and LLM first-token latency compound with TTS TTFA before the user hears anything
• Pipeline compounding example from the Coval benchmark data: if STT contributes roughly 150ms and LLM first-token roughly 250ms, a 133ms gap between two TTS providers can push a borderline pipeline over the ~800ms naturalness threshold
• P50 vs. P95 vs. P99 — why percentile framing matters more than median:
  • P50 describes the median call; P95/P99 describe the calls users complain about
  • Tail latency is where architectures are validated or disqualified in production
  • IQR (interquartile range) as a consistency proxy: a narrow IQR means the median is representative; a wide IQR means the median flatters the provider
• Transport protocol as a hidden variable: with WebSocket multiplexing, figures shift — this explains why provider A may lead on HTTP benchmarks but trail on WebSocket tests at the same infrastructure
• Throughput under concurrency as a separate failure mode: a provider hitting low TTFA on single requests may degrade under production concurrency — no published single-request benchmark captures this
• The implication for any table of numbers: carry all the figures below with their measurement provenance; vendor-reported and third-party-benchmark figures for the same model can diverge substantially

How the leading real-time TTS APIs actually perform: provider benchmarks with sources named

• Role: Delivers the roundup the reader came for — but structured around use-case routing rather than a simple latency ranking, carrying every precise figure with its source so readers can evaluate provenance.
• Framing note for writer: lead each provider entry with what it is optimised for, then give the numbers, then name the measurement source — this keeps the benchmark-provenance argument from section 2 active throughout
• Cartesia Sonic — ultra-low TTFA, SSM architecture
  • Cartesia Sonic 4: ~40ms TTFA (Cartesia-reported); Sonic-3: 90ms TTFA on Cartesia's own site
  • Third-party divergence: Coval benchmark (data captured May 4, 2026) records Sonic-3 at 188ms P50 — a gap that illustrates section 2's measurement-condition argument directly
  • Consistency caveat: 100ms IQR on Coval — wide tail, meaning the 188ms median understates worst-case production latency
  • Architecture: State Space Model (SSM) rather than standard transformers — the design choice behind its headline speed
  • Product line: Sonic-3 (TTS), Ink-Whisper (STT), Line (voice agent platform)
  • Languages: 40+ with regional accents; emotional expressiveness controls; instant voice cloning from 3 seconds of audio
  • Pricing: credit-based — Pro $4/month (100K credits, annual), Startup $39/month (1.25M credits, annual), Scale $239/month (8M credits, annual); per-character rate $0.038
  • Best for: latency-critical real-time agents, WebRTC voice bots, real-time interruption handling
• ElevenLabs Flash v2.5 — real-time multilingual conversational model
  • ElevenLabs' own claim: ~75ms latency for Flash v2.5
  • Third-party divergence: Coval benchmark places Flash v2.5 at 288ms P50 and 304ms P75 — slower than ElevenLabs Turbo v2.5 (264ms P50, 279ms P75) in the same test conditions
  • Model intended for real-time conversation; Flash v2.5 extended Flash v2 from English-only to 32 languages
  • HTTP streaming adds 40–100ms connection overhead per turn in multi-turn sessions; with WebSocket multiplexing, ElevenLabs Turbo v2.5 reached 257ms P50 in Coval conditions
  • Model lineup shift in 2026: Turbo v2.5 is deprecated on ElevenLabs' model list, Flash models are recommended over it in all use cases; the active real-time choice is now Flash v2.5
  • Pricing: Flash/Turbo models at $0.05 per 1K characters ($50 per 1M); subscription tier starts at $60 per million characters
  • Best for: emotional inflection, voice cloning nuance, dialect stability, multilingual conversational agents
• ElevenLabs Eleven v3 — expressive/batch model, not a real-time option
  • Generally available since 2 February 2026 (alpha opened June 2025)
  • 70+ languages, 380+ voices, multi-speaker dialogue, inline audio tags ([laughs], [whispering], [sarcastic])
  • Coval benchmark: Multilingual v2 records 1,232ms P50 — v3 is similarly in the non-real-time tier
  • Pricing: $0.10 per 1K characters ($100 per 1M); subscription API pricing starts at $120 per million characters
  • Best for: pre-rendered IVR prompts, audiobooks, marketing voiceovers — any audio rendered once and played many times; not for live calls
• Inworld AI Realtime TTS-2 — top-ranked real-time quality per independent arena
  • Realtime TTS-2 launched as a Research Preview in May 2026; reached general availability August 31, 2026; currently #2 on the Artificial Analysis Realtime TTS Arena with ~1,209 ELO
  • TTFB: <100ms vendor-reported; P99 ranges 130–250ms depending on model variant
  • Natural-language steering across 8 dimensions (emotion, articulation, intonation, volume, pitch, range, speed, vocal style); cross-lingual voice identity; 200+ languages
  • Three variants: Realtime TTS-2 Research Preview (newest), Realtime TTS-2 (recommended production), and Flash (lower cost)
  • Quality context: Google Studio voices score ELO 1,048, 188 ELO points below Inworld TTS-2 Research Preview on the same leaderboard
  • Best for: real-time conversational AI requiring top-tier quality without premium pricing
• Deepgram Aura-2 — enterprise voice agents, compliance-heavy verticals
  • Coval benchmark: 313ms P50; highest WER among real-time providers at 6.4%
  • Sub-200ms mean latency reported by Smallest.ai independent testing (~184ms mean)
  • Handles high concurrency with stable, consistent latency — the case for Aura-2 is throughput reliability, not headline TTFA
  • 7 languages with sub-200ms latency; domain-tuned pronunciation for numbers, names, and technical terms — healthcare, finance, compliance-critical IVR
  • No voice cloning
  • Pricing: $0.030 per 1,000 characters; $0.027 at Growth tier — the most affordable per-character rate among the real-time tier (cf. Cartesia $0.038, ElevenLabs Flash $0.05 per 1K)
  • Compliance: SOC 2 Type II, HIPAA, GDPR, CCPA, PCI DSS; on-premise deployment supported
  • Bundles with Deepgram Nova STT and LLM orchestration in Voice Agent API — relevant for teams wanting a single-vendor voice stack
  • Best for: high-volume English-language phone operations, contact centers, regulated industries
• Smallest AI Lightning v3.1 — quality-speed balance with multilingual coverage
  • TTFA: ~200ms; RTF: 0.3 (generates audio 3.3× faster than playback speed)
  • Independent blind listening tests across 1,088 samples: preferred 76.2% of the time against OpenAI GPT-4o-mini-TTS using LLM-as-a-Judge framework on Seed TTS dataset
  • 15 languages including English, Spanish, Hindi, Tamil, French, German, Italian, Portuguese, Swedish, Dutch, and Indic languages (Tamil, Telugu, Malayalam, Kannada, Marathi, Gujarati)
  • Geo-routed servers: US (Oregon) and India (Mumbai)
  • Voice cloning: instant from 5 seconds (Smallest AI's own documentation cites 5-second instant cloning; the 5–15 second / 45+ minute figures belong to Inworld AI, not Smallest AI)
  • Pricing: Pro plan $9/month at ~$0.25 per 10,000 characters pay-as-you-go; enterprise includes on-prem, 99.99% uptime SLA, HIPAA zero data retention, SOC2
  • Best for: conversational agents where audio quality and latency must coexist; multilingual deployments across English and Indic markets
• OpenAI TTS — GPT-ecosystem integration, not optimised for real-time
  • Three models: tts-1, tts-1-hd, gpt-4o-mini-tts; 50+ languages; 13 built-in voices (alloy, ash, coral, echo, fable, nova, onyx, sage, shimmer, ballad, verse, marin, cedar)
  • tts-1-hd records 2,295ms P50 — clearly outside real-time use cases
  • tts-1 is the faster variant; chunk transfer encoding enables streaming before full generation completes
  • Pricing: $15 per 1M characters
  • Best for: teams already on the OpenAI platform, batch content generation, conversational AI within an existing GPT workflow
• Google Cloud TTS — broadest language coverage, not a real-time first choice
  • 380+ voices across 75+ languages; WaveNet/Neural2 achieving 200–250ms latency in third-party benchmarks
  • Chirp3-HD voices: reported catastrophic speed drops in production (5 minutes of audio taking more than 10 minutes to generate) — a tail-latency failure mode not visible in median benchmarks
  • Free tier: 1M free characters monthly for standard voices — useful for development and testing
  • Neural TTS pricing: $16 per 1M characters
  • Best for: multilingual applications requiring breadth, GCP-native stacks; ruled out for real-time agents
• Amazon Polly — AWS-native, batch and IVR workloads
  • Generative TTS engine: 31 voices across 20 languages
  • Latency: 100ms–1 second range in third-party testing — wide variance makes it unsuitable for latency-sensitive real-time use
  • Free tier: 5 million characters/month for first 12 months
  • Pricing: Standard TTS $4/million characters; Neural TTS $16/million characters
  • Speech marks for animation sync; custom lexicons; deep AWS integration
  • Best for: AWS-native architectures, batch synthesis, IVR systems where sub-300ms TTFA is not a hard requirement
• Microsoft Azure AI Speech — widest voice and language coverage among hyperscalers
  • 400+ neural voices across 140+ languages and locales
  • Neural TTS pricing: $12.00 per 1M characters pay-as-you-go; commitment tiers drop to $9.75/1M at high volume
  • Includes Neural HD, Custom Neural Voice for branded voice training from recordings
  • October 2025: launched Voice Live API — STT + generative AI + TTS in a single low-latency interface with barge-in functionality, purpose-built for intelligent voice agents
  • Full SSML plus mstts namespace (style, role, emphasis)
  • Best for: enterprise branding, regulated industries, widest language requirement, Microsoft ecosystem integration
• Kokoro