Voice

TTS / Voice Synthesis · Series D $500M

ElevenLabs builds real-time text-to-speech and voice-cloning models with support for 29 languages and ultra-low latency.

Enterprise Voice Agents · Series D $350M

Parloa builds no-code conversational AI agents for European enterprise contact centers, specializing in voice and omnichannel automation.

Enterprise Voice Agents · Series D $250M

Decagon builds AI customer-support agents that handle ticket resolution, live chat, and voice calls for mid-market and enterprise SaaS companies.

TTS / Voice Synthesis · Series B $250M

Sesame builds fast, high-fidelity voice models for conversational AI with a focus on multilingual support and low-latency inference.

Voice Agent Infrastructure · Series C $100M

LiveKit provides open-source infrastructure for real-time audio and video, powering voice-agent orchestration and WebRTC-based communication.

Enterprise Voice Agents · Acquired NICE

Cognigy builds enterprise conversational AI for customer service, supporting voice, chat, and omnichannel deployments.

Enterprise Voice Agents · Series D $86M

PolyAI builds conversational AI for enterprise contact centers, specializing in restaurant reservations, hospitality, and logistics support.

STT / ASR · Series C $130M

Deepgram provides deep-learning-based automatic speech recognition and text-to-speech APIs for real-time transcription and voice synthesis.

Cartesia develops ultra-low-latency voice models including Sonic, achieving ~40ms time-to-first-audio for real-time conversational AI.

TTS / Voice Synthesis · Series B $100M

Voice Agent Infrastructure · Series B $50M

Vapi provides a developer platform for building and deploying voice AI agents with sub-second latency and flexible integrations.

Voice Dictation · Series A $25M

Wispr Flow makes an AI voice-dictation app that turns natural speech into clear, formatted text in any application across Mac, Windows, iPhone, and Android.

Resemble AI provides voice cloning and custom text-to-speech APIs for gaming, entertainment, and call-center applications.

TTS / Voice Synthesis · Series B $13M

Speech Translation · Series C $300M

DeepL provides neural machine translation and launched DeepL Voice in 2024 for real-time speech-to-speech translation.

Dialpad provides a cloud business communications platform with built-in voice AI for real-time transcription, sentiment analysis, and agent coaching.

Contact Center AI · Series F $170M

Cresta provides real-time AI coaching and automation for contact-center agents, analyzing voice and chat to improve sales and support outcomes.

Contact Center AI · Series D $125M

Kore.ai provides an enterprise conversational AI platform for contact centers, covering voice, chat, and automation across 100+ channels.

Enterprise Voice Agents · Growth $150M

Contact Center AI · Series C $125M

Observe.AI provides voice analytics and agent-assist tools for contact centers, analyzing 100% of calls for compliance and coaching.

SoundHound AI builds a voice-native conversational AI platform powering in-car assistants, restaurant drive-thru and phone ordering, and enterprise agents through its Amelia platform.

Enterprise Voice Agents · Market cap $2.8B

STT / ASR · Series C $50M

AssemblyAI provides AI-powered speech-to-text APIs with sentiment analysis, speaker diarization, and real-time transcription features.

Voiceflow provides a collaborative platform for designing, prototyping, and deploying conversational AI agents across voice and chat.

Voice Agent Infrastructure · Series B $15M

Accent Conversion · Series B $65M

Sanas builds real-time accent conversion technology for contact-center agents, translating non-native accents to standard dialects on live calls.

Yellow.ai provides a conversational AI platform for customer service and employee support with voice, chat, and automation features.

Enterprise Voice Agents · Series C $78.2M

Emotion & Prosody Models · Series B $50M

Hume AI builds empathic voice models that detect and generate emotional prosody in speech for more human-like conversational agents.

Emotion & Prosody Models

Bland AI offers a platform for deploying outbound and inbound phone-based conversational agents at scale.

Voice Agent Infrastructure · Series B $40M

Speech Translation · Series A $21M

KUDO offers a cloud-based platform for multilingual meetings and events with AI-assisted interpretation and real-time translation.

Synthflow offers a no-code platform for building voice AI agents, targeting SMBs and solo entrepreneurs without technical resources.

Voice Agent Infrastructure · Series A $20M

STT / ASR · Series B $62M

Speechmatics offers automatic speech recognition in 50+ languages with real-time transcription and on-premises deployment options.

WellSaid Labs creates synthetic voice avatars for corporate training, e-learning, and media production with studio-quality output.

TTS / Voice Synthesis · Series A $10M

Enterprise Voice Agents · Series C $30M

Rasa builds open-source and enterprise conversational AI for on-premise voice and text assistants, popular in regulated industries.

TTS / Voice Synthesis · Series A $10M

Murf AI provides a studio-quality text-to-speech platform with 120+ AI voices for video production, e-learning, and marketing content.

AI Personal Assistants · Seed $10.7M

Lindy builds AI personal assistants that handle email, scheduling, and voice calls for individuals and small teams.

AI Personal Assistants

Rime develops fast, expressive text-to-speech models optimized for conversational AI and live-agent experiences.

TTS / Voice Synthesis · Seed $5.5M

Voice Agent Infrastructure · Seed $4.6M

Retell AI provides APIs and SDKs for building conversational voice agents with sub-800ms latency and built-in telephony support.

Vertical Voice Agents · Seed $4M

Goodcall provides AI phone answering services for small businesses, serving 42,000+ barbershops, salons, and local service providers.

Vertical Voice Agents

Phonic provides conversational voice surveys and qualitative research tools powered by AI transcription and sentiment analysis.

Voice Surveys & Insights · Seed $4M

Voice Surveys & Insights

Dasha builds conversational AI for voice and text, offering a low-code SDK for developers building phone-based agents.

Voice Agent Infrastructure · Seed $2M

Vertical Voice Agents · Pre-Seed $500K

Arini builds AI phone agents for dental practices, automating appointment scheduling, insurance verification, and patient follow-up.

Vertical Voice Agents

Fish Audio builds multilingual TTS and voice-cloning models with support for Chinese, English, and other Asian languages.

—

HQ

Soniox

Soniox develops high-accuracy, low-latency speech recognition optimized for conversational AI and telephony applications.

STT / ASR

—

HQ

No roster changes this week. The roster is re-evaluated quarterly; read the methodology in the disclosures.

As of 2026-07-25

Benchmarks

See all 5 →

Word Error Rate

% WER

Transcription errors across the Open ASR test suite — lower is better.

↓ lower is betterOpen ASR Leaderboard

#01
modulate/vfast
3.1%
#02
HojoAI/Hojo-ASR-V1
3.1%

KPIs

Total funding YTD

$0

across 7 rounds

Companies tracked

0

5 countries

Largest round YTD

$0

Sierra · Series D

Median round YTD

$0

across 7 rounds

News volume · 7d

0

-15% vs prior 7d

Signals

Venture capital deployed · 2026 YTD

$2.3B

7 announced rounds

vs $1.5B in 2025

Sector market cap

$2.8B

1 public company

largest = 100% of sector

Largest raise · trailing 12mo

$950M

Series D · Sierra

Catalysts ahead · next 12mo

9

3 conferences · 1 earnings · 3 launches

Latest News

View all 100+ →

Videos

View all 760 →

Talent Moves

May 12, 2026
Morgan Norman
Chief Marketing Officer
From
Chief Marketing Officer
at Dialpad
To
Chief Marketing Officer
at PolyAI
PR Newswire
May 12, 2026
Danielle Waknine
VP, Go-to-Market Vertical Expansion
From
VP, B2B Vertical
at Shopify
To
VP, Go-to-Market Vertical Expansion
at PolyAI
PR Newswire
May 12, 2026
Helen Greul
SVP, Product Engineering
From
Prior engineering leadership role
at Undisclosed
To
SVP, Product Engineering
at PolyAI
PR Newswire
Jan 22, 2026
Alan Cowen
CEO & Founder
From
CEO & Founder
at Hume AI
To
Research executive (Gemini voice)
at Google DeepMind
TechCrunch
Jan 20, 2026
Karthik Rajaram
General Manager & Country Head, India
From
General Manager & VP, India and South Asia
at Elastic
To
General Manager & Country Head, India
at ElevenLabs
Business Today
Jan 13, 2026
OfOne founding team
Engineering & product team
From
Founders
at OfOne
To
Acquired — Deepgram for Restaurants product line
at Deepgram
Deepgram
Sep 8, 2025
Cognigy leadership team
Executive team
From
Independent leadership
at Cognigy
To
Acquired ($955 M)
at NICE Ltd.
SEC filing (NICE)
Sep 1, 2023
Brendan Iribe
Co-founder & CEO
From
Co-founder & former CEO, Oculus VR
at Oculus / Meta
To
Co-founder & CEO
at Sesame
TechCrunch
May 19, 2023
Ping Wu
CEO
From
VP, Conversational AI Products (Dialogflow, Speech, Translation)
at Google
To
CEO
at Cresta
PR Newswire
May 19, 2023
Zayd Enam
Co-founder & CEO
From
Co-founder & CEO
at Cresta
To
Co-founder, Advisory Role
at Cresta
Cresta Blog
Feb 1, 2023
Bret Taylor
Co-founder & CEO
From
Co-CEO
at Salesforce
To
Co-founder & CEO
at Sierra
Fortune

Catalysts

Conferences

Major industry dates · soonest first

Earnings Calls

Public roster companies · forecast from SEC filings

Venture Stages

1

Seed

2

Series A

3

Series B

4

Series C

5

Series D

6

Public & Acquired

Valuations

Sized by valuation · area is proportional to market value

Public caps as of Jul 23, 2026

Public · real-time market capPrivate · last-round valuation

2 companies without valuation on file

Funding & analysis

Company

18

19

20

21

22

23

24

25

26

Total Raised

Bottlenecks

Multiple bets

Sub-200ms end-to-end conversational latency

Human conversation tolerates pauses of roughly 200–300ms; beyond that, speakers perceive the system as unresponsive. A cascaded pipeline (ASR → LLM → TTS + network) compounds latency at every stage, routinely landing at 800ms–1.5s in production. Shrinking the entire round-trip below 200ms while preserving model quality requires rethinking each component and how they are pipelined together, not merely optimizing each piece in isolation.

Approaches in flight

Sub-200ms end-to-end conversational latency

▸ Streaming-first cascade with sub-token TTS

Cartesia's Sonic 3.5 (~40ms TTFA) and Deepgram Aura-2 (~90ms) reach first audio before the LLM finishes generating, by streaming token-by-token into TTS. Vapi and Retell AI build orchestration layers on top of these primitives to shave further milliseconds from infrastructure overhead.

▸ Native speech-to-speech (end-to-end) models

GPT-4o Voice and Gemini Live bypass the text intermediate by passing audio directly through a single multimodal model. This removes TTS/ASR handoff latency but introduces new trade-offs in controllability and cost; Hume AI's EVI 4 mini pursues the same architecture for the emotional fidelity use case.

▸ On-device / edge inference

Running ASR and small TTS models locally eliminates round-trip network time entirely. Soniox and Speechmatics offer on-prem deployments; Apple and Google push ambient voice processing directly on device, which matters for wearables and always-on assistants.

Decade away

Robust turn-taking and endpointing in noisy environments

Detecting when a human has finished speaking — and distinguishing a mid-sentence pause from a true turn-end — is harder than it looks. Premature endpointing causes the agent to interrupt; late endpointing adds hundreds of milliseconds of dead air every turn. Background noise, fast speech, heavy accents, and emotional speech all degrade endpointing models trained on clean studio data.

Approaches in flight

Robust turn-taking and endpointing in noisy environments

▸ Acoustic end-of-utterance classifiers

Deepgram and AssemblyAI ship proprietary VAD and endpointing models trained on production call audio. Goodcall and Retell AI expose tunable silence-threshold parameters so builders can calibrate for their specific acoustic environments.

▸ Semantic + prosodic joint modeling

Rather than relying purely on silence, these models use semantic completeness (did the sentence resolve?) combined with prosodic cues (falling intonation). Hume AI's emotion-aware models and Speechmatics' prosody-aware pipeline incorporate both signals, though robust real-world performance remains a research challenge.

Approach emerging

High-fidelity zero-shot voice cloning from short samples

Voice cloning from a few seconds of audio has improved dramatically, but replicating the full acoustic identity of a speaker — including their characteristic prosody, emotional range, breathiness, and accent — without artifacts still falls short of human indistinguishability for many voices and languages. The problem is especially hard for voices underrepresented in training data.

Approaches in flight

High-fidelity zero-shot voice cloning from short samples

▸ Diffusion-based and flow-matching TTS

ElevenLabs' Eleven v3 (GA February 2026) and PlayAI use flow-matching architectures that model fine-grained acoustic detail, enabling emotional control via audio tags and dramatically reducing cloning artifacts at 70+ language coverage.

▸ Codec-based neural voice models

Cartesia and Fish Audio encode voices as compressed neural codec tokens, enabling very fast inference at competitive quality. Resemble AI's approach layers on perceptual loss training to close the remaining gap on naturalness metrics.

▸ Retrieval-augmented voice personalization

WellSaid Labs and Murf AI focus on high-fidelity studio-quality voices with full persona control rather than arbitrary zero-shot cloning, targeting enterprise branding use cases where repeatability matters more than novelty.

Decade away

Real-time code-switching and multilingual ASR

In real calls, speakers switch languages mid-sentence, mix dialects with their base language, and use domain-specific terminology in non-English markets. Current ASR models excel on their dominant training language but accuracy collapses significantly for accented speakers, low-resource languages, and intra-sentence code-switching. Routing overhead (detect language → switch model) adds latency and fails at mixing points.

Approaches in flight

Real-time code-switching and multilingual ASR

▸ Single-model multilingual ASR

Deepgram Nova-3 Multilingual supports real-time transcription in 10+ languages simultaneously without routing. Speechmatics covers 55 languages with streaming as a first-class offering and consistently outperforms on accented English in independent benchmarks.

▸ Universal speech encoders

Whisper-derived and internally trained universal encoders (Google Chirp, AssemblyAI Universal-2) attempt broad language coverage through massive multilingual training corpora. Coverage breadth still trades against per-language accuracy at the frontier.

▸ Real-time translation layer

Cresta and KUDO address the multilingual contact-center problem differently — translating agent speech in real time rather than requiring multilingual ASR. Cresta's Real-Time Translation now covers 30+ languages across voice and chat.

Decade away

Detecting and responding to emotional and paralinguistic signals

A large fraction of conversational meaning is carried not in words but in tone, pace, stress, hesitation, and emotional coloring. Current voice agents are largely blind to these signals: they read frustration the same as contentment, miss sarcasm, and respond to distressed callers with the same register as relaxed ones. Building models that reliably infer and act on paralinguistic state — across acoustic diversity — is an open problem.

Approaches in flight

Detecting and responding to emotional and paralinguistic signals

▸ Emotion-native voice models

Hume AI's Empathic Voice Interface (EVI) is purpose-built around the premise that emotion is a first-class signal: the model tracks continuous valence and arousal and modulates its own voice output accordingly. Octave 2 launched with sub-200ms latency and multi-speaker conversation support.

▸ Sentiment-enriched ASR streams

AssemblyAI and Observe.ai offer sentiment and emotion detection as an optional enrichment layer on top of transcription, enabling contact-center supervisors to surface frustrated calls in real time. Sanas focuses specifically on reducing accent-driven emotional misread by normalizing acoustic features before downstream interpretation.

Decade away

Persistent cross-session memory for voice agents

A phone agent that remembers nothing from previous calls forces customers to re-authenticate and re-explain context on every call — destroying the UX advantage of AI. Storing and retrieving conversation history at sub-second latency while respecting privacy regulations, avoiding hallucinated recall, and scaling to millions of simultaneous users is an unsolved systems problem with no agreed architecture.

Approaches in flight

Persistent cross-session memory for voice agents

▸ Vector-based episodic memory

Sierra and PolyAI embed conversation summaries and customer facts into vector stores, retrieving relevant context before each turn. This works for structured recall but degrades on nuanced behavioral history and long time horizons.

▸ CRM-grounded identity

Cognigy, Kore.ai, and enterprise contact-center platforms anchor agent memory to existing CRM records, making it an integration problem as much as an AI problem. The challenge is keeping CRM data current enough to be useful in real-time conversation.

Decade away

Reliable voice deepfake detection and trust infrastructure

Zero-shot voice cloning has crossed the threshold of human indistinguishability for many voices; impostor scam losses grew 45%+ from 2023 to 2025. No deployed detection system achieves reliable real-time liveness detection across the diversity of phones, codecs, languages, and cloning architectures in the wild. STIR/SHAKEN caller-ID infrastructure helps with spoofed numbers but does nothing about the voice itself.

Approaches in flight

Reliable voice deepfake detection and trust infrastructure

▸ Acoustic deepfake classifiers

Resemble AI ships Resemble Detect as a standalone API for real-time audio authenticity classification. OmniSpeech and several stealth-mode anti-fraud startups target enterprise and telecom deployments. All current classifiers degrade on unseen cloning architectures.

▸ Provenance watermarking

ElevenLabs includes inaudible watermarks in generated audio as a content-provenance signal; AudioSeal (Meta Research) proposes a standardized approach. The limitation is that watermarks can be destroyed by transcoding or compression, and adversarial actors simply disable them.

▸ Regulatory and carrier-level authentication

The FCC's 2024 ruling classifying AI voice cloning in robocalls as an artificial voice under TCPA opened enforcement pathways. Proposed 2026 regulations require separate consent for AI-generated calls. Telecom carriers are rolling out updated STIR/SHAKEN protocols as a complementary signal.

Investment Theses

Thesis

Voice AI eliminates the $340B contact-center labor pool

Global contact centers spend roughly $340B per year on human agents, most of whom handle repetitive tier-1 calls resolving the same small set of intents. At cents per minute versus $29–42/hour for a U.S. agent, AI voice handles those calls at a fraction of the cost with no attrition — and the market is now paying up for the wager: Sierra raised $950M at $15.8B in May 2026 on $200M ARR, Parloa tripled to $3B in eight months, and Decagon tripled to $4.5B. Gartner's April 2026 survey found 31% of contact centers planning frontline workforce reductions through Q1 2027 and 63% already cutting agent headcount via attrition. The thesis: the winning platforms capture enterprise contracts worth 10–30% of the labor they displace, building a new category of high-margin software on top of commodity model APIs.

Counter-thesis

Voice AI eliminates the $340B contact-center labor pool

Contact-center work is far less homogeneous than the thesis assumes once you move below tier-1, and enterprise contracts demand CRM integration, compliance coverage, and human-in-the-loop workflows that erode margins. Incumbents (Salesforce, ServiceNow, Twilio) are adding AI voice natively, and with Sierra priced at ~79x ARR the platforms are already valued as if displacement is complete — any slip in seat-replacement economics compresses the whole category.

Thesis

Developer infrastructure becomes the dominant value capture layer

In every major platform cycle, picks-and-shovels infrastructure accumulates durable value even as the application layer commoditizes. Voice is repeating the pattern, and 2026 has priced it in: LiveKit — the real-time engine behind ChatGPT's voice mode — hit $1B in January carrying 2.5B calls a year; Deepgram raised $130M at $1.3B with Twilio, ServiceNow, SAP and Citi as strategic backers; Vapi cleared $500M after Amazon Ring picked it over 40 rivals and crossed 1 billion cumulative calls. The thesis: the orchestration abstraction — BYO-model, SLA guarantees, HIPAA-eligible compliance rails, and now agent-ops tooling like Retell's Conductor — is stickier than any individual application and commands usage-based revenue that scales with the entire market.

Counter-thesis

Developer infrastructure becomes the dominant value capture layer

Infrastructure is only defensible with a genuine technical moat or deep switching cost. Cloud providers are shipping competing primitives at subsidized prices, OpenAI and Google keep folding realtime voice into their model APIs, and Deepgram's own strategic round shows incumbents buying optionality rather than ceding the layer. The middle of the stack risks becoming a pass-through margin business squeezed from below by model APIs and from above by verticalizing applications.

Thesis

Emotion-aware voice is the next defensible model layer

Current voice agents are acoustically competent but emotionally illiterate — they miss frustration, sarcasm, distress, and engagement signals humans catch instantly. The bet: building emotion detection and response into the model layer creates a qualitatively different product that can't be replicated by bolting sentiment analysis onto a generic TTS pipeline. Google just validated the wager at the highest level — DeepMind acqui-hired Hume AI founder Alan Cowen and his core prosody team in January 2026 and licensed the EVI technology to make Gemini an empathic companion. The refined thesis: emotional intelligence in voice is real, scarce, and worth paying for; the open question is whether independents can compound on it before the foundation labs absorb it.

Counter-thesis

Emotion-aware voice is the next defensible model layer

The Hume deal cuts both ways: Google paid for the team and a license, not the company — evidence that emotion-aware voice is a feature the foundation labs can hire their way into rather than a durable standalone business. Emotion detection also remains noisy and culturally variable, and companionship use cases face regulatory and reputational scrutiny that could cap the addressable market.

Thesis

Vertical-specific voice agents take their categories before horizontal platforms lock in

Dental practices, legal intake, SMB answering, healthcare triage, and restaurant reservations all share the same insight: 80% of calls resolve the same five intents, and a voice agent trained on industry-specific data with deep vertical integrations outperforms a generic agent for those intents. The thesis: purpose-built vertical winners like Arini (dental), Goodcall (SMB answering at scale), and Lindy will lock up their categories before horizontal platforms have time to verticalize, because switching costs compound once a voice agent is wired into practice-management or POS software — healthcare call automation alone is now a named budget line as hospitals chase hold-time and burnout reductions.

Counter-thesis

Vertical-specific voice agents take their categories before horizontal platforms lock in

Vertical depth is a two-edged sword: it creates switching costs but caps TAM and makes expansion expensive. Horizontal platforms flush with 2026 mega-rounds — Sierra's $950M, Parloa's $350M — are explicitly funding vertical modules, and no-code builders like Voiceflow let any systems integrator spin up a vertical product without the startup's headcount. Category lock-in may arrive more slowly than the vertical thesis assumes.

Thesis

Real-time multilingual voice unlocks a global B2C customer base

English-only voice AI serves roughly 20% of the world's population at native quality; the next TAM expansion requires robust multilingual ASR, TTS, and translation at conversation latency, which is only now viable. ElevenLabs — $500M ARR in May 2026 and courting a $22B tender valuation — is executing exactly this play: 70+ language TTS, a Japan contact-center push with NTT Docomo Business, a Poland ecosystem deal, and a stated shift to 70% enterprise revenue by end-2027. Deepgram's Nova-3 does simultaneous real-time transcription across 10+ languages and Cartesia calls Sonic-3.5's multilingual gains its headline feature. Companies that solve 30+ language quality reposition their TAM from U.S. enterprise to global e-commerce, financial services, and government — a 5–10x expansion.

Counter-thesis

Real-time multilingual voice unlocks a global B2C customer base

Multilingual quality degrades non-linearly in low-resource languages, and cultural localization — register, idioms, turn-taking norms — requires human editorial work that erodes gross margin. The global push also walks straight into the voice-cloning backlash: fraud epidemics from Canada to Korea, a UK legal gap the BBC says current law may not cover, and a pending NO FAKES Act mean every new market adds compliance surface that advantages established local players.

Top 10

Investors

By tracked rounds led

Books

Relevancy
Most recent

Graveyard

Air.ai
FTC filed suit in 2025 alleging deceptive marketing of business opportunities and false earnings claims. March 2026 settlement banned the owners from marketing business opportunities, imposed an $18M monetary judgment (largely suspended), and effectively put the company out of business. Platform was inactive by 2025 with key features broken and refunds stalled.
Obituary

#03

elevenlabs/scribe_v2

3.3%

#04

Qwen/Qwen3-ASR-1.7B-hf

3.5%

#05

assemblyai/universal-3-5-pro

3.5%

#06

soundsgoodai/Zipformer-cr-ctc-transducer-XL-290M

3.7%

#07

gladia/solaria-3

3.7%

#08

Qwen/Qwen3-ASR-0.6B-hf

4.0%

#09

nvidia/nemotron-speech-streaming-en-0.6b

4.0%

#10

AutoArk-AI/Audio8-ASR-0.1B

4.9%

#11

microsoft/azure-speech-06-2026

5.2%

#12

OpenMOSS-Team/MOSS-Transcribe-preview-2B

5.3%

Throughput (RTFx)

× real-time

Audio-seconds transcribed per second of compute — higher is faster.

↑ higher is betterOpen ASR Leaderboard

#01
nvidia/stt_en_conformer_ctc_small
6,258×
#02
nvidia/parakeet-tdt_ctc-110m
6,119×
#03
nvidia/parakeet-tdt-0.6b-v3
6,098×
#04
nvidia/parakeet-tdt-0.6b-v2
6,038×
#05
nvidia/stt_en_fastconformer_transducer_large
6,033×
#06
nvidia/stt_en_fastconformer_ctc_large
6,027×
#07
nvidia/parakeet-ctc-0.6b
5,884×
#08
nvidia/stt_en_conformer_ctc_large
5,732×
#09
nvidia/stt_en_conformer_transducer_small
5,557×
#10
nvidia/parakeet-rnnt-0.6b
5,407×
#11
nvidia/parakeet-ctc-1.1b
5,015×
#12
nvidia/parakeet-tdt-1.1b
4,529×

TTS naturalness

Elo

Crowd-voted naturalness of synthesized speech (TTS Arena).

↑ higher is betterTTS Arena

#01
Vocu V3.0
1,565
#02
CastleFlow v1.0
1,562
#03
Inworld TTS MAX
1,560
#04
Papla P1
1,551
#05
Luna TTS
1,550
#06
Inworld TTS
1,544
#07
Luck Dolphin Turbo
1,542
#08
Luck Dolphin
1,542
#09
star-june-2026
1,536
#10
MiniMax Speech 2.8 HD
1,534
#11
Gradium TTS
1,532
#12
Hume Octave
1,531

Transcription cost

$ / minute

Published speech-to-text API price per minute — lower is cheaper.

↓ lower is betterLiteLLM (open pricing)

#01
Groq Whisper v3 Turbo
$0.0007
#02
AssemblyAI Best
$0.0020
#03
ElevenLabs Scribe
$0.0037
#04
Deepgram Nova-3
$0.0043
#05
OpenAI Whisper
$0.0060
#06
Azure Whisper
$0.0060

TTS cost

$ / 1M chars

Published text-to-speech price per million characters — lower is cheaper.

↓ lower is betterLiteLLM (open pricing)

#01
AWS Polly Standard
$4
#02
OpenAI TTS
$15
#03
AWS Polly Neural
$16
#04
OpenAI TTS HD
$30
#05
Google Chirp
$30
#06
Groq PlayAI
$50
#07
MiniMax Speech Turbo
$60
#08
ElevenLabs v3
$180

Sentiment · 7d

0

Positive

0

Neutral

0

Negative

Hot company · 7d

Deepgram

4 mentions

Most patents filed · 12mo

01Sesame42
02SoundHound AI13
03Sierra8

Most raised · all-time

01Sierra$1.6B
02ElevenLabs$781M
03Parloa$557M

Senior talent moves · 18mo

11

founder + C-suite moves tracked

Open frontier bottlenecks

7

mapped in Hard Problems

Companies tracked

40

1 public · 39 private

positive

OpenAI makes its Presence felt in CX

OpenAI launches a platform for enterprises to deploy and manage AI agents across workflows.

No Jitter

2dResearchneutral

Integration and governance prove critical to AI agent rollouts

Forrester-Boomi study finds organizations prioritizing integration and governance see more successful AI agent deployments.

No Jitter

Jul 14Launchpositive

Zoom taps into big market for AI receptionists

Zoom opens its AI receptionist to any business phone system, broadening addressable market.

No Jitter

Jul 14Researchneutral

The Download: Claude’s inner workings, and the future of world models

Anthropic's latest research reveals a new technique for inspecting its models' internal reasoning paths.

MIT Technology Review

Jul 14Researchneutral

Claude responds with more warmth in Hindi and more rigor in Russian, showing how language shapes AI answers

Anthropic study maps value concepts across languages in Claude, finding systematic differences in tone by language but raising methodological questions.

The Decoder

Jul 4Othernegative

MBC Announcers Confront AI Voice Cloning Crisis

MBC announcers report unauthorized AI voice cloning, raising concerns over voice identity theft and media deepfakes.

조선일보

ElevenLabs targets $22bn valuation with fresh share sale, reports say

ElevenLabs targets a $22B valuation in a fresh share sale.

Sifted

ElevenLabs at $22B: How a Voice AI Company Doubled Its Valuation in Five Months

ElevenLabs' valuation hits $22B, doubling in five months.

FourWeekMBA

Jul 3Othernegative

AI voice-cloning scams are fooling Canadian parents — but there were red flags that gave away one scammer

AI voice-cloning technology enables scams targeting parents via fake emergency calls impersonating their children.

Yahoo! Finance Canada

ElevenLabs plans share sale in September at USD 22 bln valuation - report

ElevenLabs plans a secondary share sale in September at a $22B valuation.

Telecompaper