Skip to content

Speech and Audio Models: Four Years from Whisper Rewriting Open ASR to ElevenLabs Consolidating Voice APIs

Aug 26, 2026 1 min
TL;DR Speech models split into two lines. Whisper led ASR: its MIT-licensed 1.55B model drove transcription cost toward zero in September 2022; after v2, v3, and turbo cut the decoder from 32 layers to four, OpenAI moved to closed gpt-4o-transcribe. In TTS, ElevenLabs grew to an $11B valuation and $500M ARR, while Kokoro (82M, Apache-2.0) and Chatterbox preserved self-hosting. Speech-to-speech Realtime APIs are now rewriting live conversation.
Table of Contents
  1. ASR: Whisper's Evolution and Turning Point
  2. TTS: ElevenLabs' Consolidation and the Open Response
  3. Architecture: Three Decisive Ideas
  4. Licensing and Pricing
  5. Selection for Three Agent Scenarios
  6. References

🌏 中文版

In September 2022, OpenAI released Whisper on GitHub under MIT: an encoder-decoder trained with weak supervision on 680K hours of web audio, up to 1.55B parameters. Commercial ASR was expensive per hour; Whisper made near-SOTA transcription runnable on a laptop and rewrote the market. On the other branch, ElevenLabs was founded in 2022 and four years later became the de facto voice API leader after a $500M Series D at an $11B valuation and more than $500M ARR. This fourteenth family article separates speech into ASR and TTS.

For benchmark interpretation, see the AI Model Evaluation Sources Guide. This is the fourteenth family deep dive in the AI Model Landscape Overview.

ASR: Whisper's Evolution and Turning Point

VersionReleaseParametersKey facts
Original Whisper2022-0939M–1550MSix sizes plus four English-only variants; paper, GitHub, MIT
large-v22022-121550MSame architecture retrained, about 10–20% lower error (discussion)
large-v32023-111550MMel bins 80→128, new tokenizer, broad multilingual WER improvements
large-v3-turbo2024-09-30~800MDecoder cut 32→4 layers, about 8× faster, near large-v2 quality, CLI default; larger Thai/Cantonese regressions
Open line stalls2024–presentNo open Whisper weights after turbo
gpt-4o-transcribe / mini2025-03-20UndisclosedClosed API; OpenAI claims 2.46% English FLEURS WER and fewer hallucinations; explicitly not open

Two turning points matter. Turbo closes the open chapter. Inspired by Distil-Whisper, it prunes and fine-tunes large-v3: community VRAM estimates fall from about 10GB to 6GB and speed improves eightfold, but translation disappears and some languages regress. OpenAI later called closed gpt-4o-transcribe much larger than Whisper and unsuitable for local use. NVIDIA Parakeet and Mistral Voxtral inherited the open frontier; AA-WER v2.0 places Parakeet near the front.

ASR now costs below half a cent per minute. Current OpenAI prices are $0.0045/minute for gpt-transcribe, $0.006 for gpt-4o-transcribe, and $0.003 for mini. ElevenLabs batch Scribe v2 costs about $0.22/hour. Transcription is a commodity; competition moved to diarization, structured timestamps, and streaming.

TTS: ElevenLabs' Consolidation and the Open Response

MilestoneDateKey facts
ElevenLabs founded2022Two Polish founders, starting with dubbing
Multilingual v2202329-language high-quality synthesis; long-form narration workhorse
Scribe enters ASR2025-02-2699 languages, diarization, word timestamps; independent v1 WER 7.7%
Eleven v3 alphaMid-2025Audio tags such as [laughs], multi-speaker dialogue
Flash v2.52024-12~75ms, 32 languages, live-agent default; Turbo deprecated
Eleven v3 GA2026-02-0270+ languages, but not real-time, 5,000 characters/request
Series D2026-02-04$500M at $11B, led by Sequoia; $781M total funding
ARR exceeds $500M2026-05NVentures and BlackRock join extension; May 7 price cuts, TTS −55%

ElevenLabs sells a suite: v3 for expression and scriptable stage directions, Multilingual v2 for long form, Flash v2.5 for real time with a 40,000-character limit—eight times v3—Scribe for ASR, plus agents, effects, and music. It is not unbeatable: Artificial Analysis places v3 around 14th on Provider Voice while it remains among the most expensive at $100/million characters. The premium buys workflow integration, not an unassailable quality gap.

ModelParametersLicenseCloningPositioning
Kokoro-82M82MApache-2.054 voices/8 languages, ~86MB quantized, browser-capable, faster than real time on ordinary CPUs but below real time on Raspberry Pi-class boards; ~$1,000 training cost
XTTS v2~467MCPML noncommercial✓ (~6s)Former zero-shot leader; Coqui closed in 2024 and license page is gone; idiap fork
Chatterbox0.5BMIT✓ (~5s)Practical commercial cloning, emotion exaggeration control
Fish Speech~500MResearch License, noncommercial✓ (10–30s)Multilingual cloning; newer releases prohibit commercial use
Meta SeamlessM4T-v2CC-BY-NC 4.0Speech translation, FLEURS WER 18.5% across 77 languages; Voicebox weights never released

Kokoro showed that a few hundred hours of clean licensed data, 82M parameters, and about $1,000 of GPU time can produce commercial narration. It cannot clone, but makes offline narration nearly free. Chatterbox is the practical 2026 commercial cloning choice; Fish Speech is now noncommercial and XTTS's technical legacy survives without a viable commercial license.

Architecture: Three Decisive Ideas

Encoder-decoder plus weak supervision. Whisper processes 30-second log-Mel spectrograms and autoregressively emits tokens. Its advantage was 680K hours of weakly labeled multilingual audio, robust to noise, accents, and domains, with one model handling recognition, translation, and language ID. Turbo shows much decoder capacity was redundant.

Zero-shot cloning uses an audio prompt. XTTS, Chatterbox, and Fish Speech compress seconds of reference audio into a speaker embedding. Newer models use LLM backbones—Chatterbox uses Llama, Orpheus Llama 3—to autoregress discrete audio tokens and inherit LLM inference infrastructure. ElevenLabs Professional Cloning trains on tens of minutes of material for a higher ceiling.

Live conversation is moving from cascades to speech-to-speech. ASR→LLM→TTS cascades add delay and lose emotion. OpenAI Realtime reached GA in August 2025; GPT-Realtime-2 in May 2026 adds adjustable reasoning, native SIP, 60-minute sessions, and $32/$64 per million audio tokens—an estimated $0.23/minute that varies with usage. Gemini Live costs $3/$12, about $0.036/minute estimated, but recomputes full context each turn, needs compression and reconnection around 15 minutes, and requires Twilio/LiveKit for SIP. ElevenLabs stays with cascades through 75ms Flash v2.5 plus Agents. Use OpenAI for telephony and long sessions, Gemini for consumer volume, and ElevenLabs when voice identity and tags are the product.

Licensing and Pricing

ProviderModelLicenseApproximate priceBest for
OpenAI Whisperlarge-v3 / turboMIT, self-hostableYour GPUPrivate/local transcription
OpenAI APIgpt-4o-transcribeClosed$0.006/minuteAccurate batch transcription
OpenAI RealtimeGPT-Realtime-2ClosedAudio $32/$64 per 1M tokensPhone agents
ElevenLabs TTSv3 / Flash v2.5Closed$0.10 / $0.05 per 1K charactersDubbing and agent voices
ElevenLabs ScribeScribe v2ClosedBatch $0.22/hourStructured high-accuracy transcripts
Kokoro-82Mv1.0Apache-2.0Free, CPU-capableOffline narration and edge devices
ChatterboxV3MITFree, GPU recommendedCommercial voice cloning
Meta SeamlessM4T-v2CC-BY-NCFree researchSpeech translation research

Two traps: XTTS v2's CPML has no commercial licensing path after Coqui's closure. Models such as F5-TTS and MaskGCT may have MIT code but CC-BY-NC weights. Fish Speech moved from a permissive reputation to a Research License. Always inspect the current weight license, not the repository license or an old impression.

Selection for Three Agent Scenarios

Batch transcription: self-host Whisper large-v3-turbo for privacy or cost. For maximum accuracy and diarization, choose Scribe v2 or gpt-4o-transcribe; both cost cents per hour and differ mainly in structured output. Turbo has known language regressions, so test Chinese audio with your own WER set.

Dubbing and content: use ElevenLabs v3 for its unique directorial audio tags, Kokoro for high-volume offline narration at zero marginal cost, and Chatterbox or ElevenLabs Instant/Professional Cloning for commercial voice identity.

Live voice: OpenAI Realtime for PSTN with native SIP and 60-minute sessions; Gemini Live for large app/browser volume at over six times lower unit cost, while handling reconnection yourself; Flash v2.5 + Agents when a branded voice is the product. A Whisper + LLM + Kokoro cascade remains the cheapest fully self-hosted route if you engineer the latency.

In four years, Whisper made listening free infrastructure, ElevenLabs turned speaking into a $500M-ARR business, and the next contest moved to speech-to-speech: whoever fits latency inside the natural rhythm of conversation owns the next interaction layer.


References