Step 14 of 22
Speech & voice models
Transcribe speech and generate natural voices. Voice applications live or die on latency and accuracy with accents and noise. Consent rules matter for cloned voices.
How to choose
- Live voice agents: low-latency streaming models on both ends.
- Batch transcription: accuracy and price per hour.
- Voice cloning: only with consent, and from a provider with misuse safeguards.
What to look for
The questions that separate the options, in order of how much they matter.
01
Speech to text: word error rate on a stated test set
Word error rate (WER) is the share of words a model gets wrong, so lower is better, but it only means something with the test set named. Artificial Analysis's AA-WER v2 averages about 8 hours of audio: agent-style conversations (50%), European Parliament speech (25%) and earnings calls (25%), with cleaned reference transcripts. In October 2026 the leaders sit around 2%, while Whisper large-v3-turbo scores 4.6%. Vendor WER claims on their own data are not comparable; test on audio that sounds like yours, with your accents, jargon and noise.
02
Speech to text: speed, streaming and features
Batch transcription is judged by speed factor, seconds of audio per second of processing; Deepgram Nova-3 runs over 600 times real time, Scribe v2 about 63. Live captions and voice agents need a true streaming model with low latency, which is often a separate model and price (Scribe v2 Realtime, Gemini 3.5 Transcribe Live, gpt-live-transcribe). Then check diarization (who spoke), word timestamps, custom vocabulary and language detection, and note that some features cannot be combined in one request.
03
Text to speech: blind-vote quality, read with care
Artificial Analysis's Speech Arena ranks voices by Elo from listener preferences, with each model using its own built-in voices; a separate board uses controlled voices. Eleven v4 and v4 Turbo lead near 1330, with Cartesia Sonic 3.6 and Gemini 3.8 Flash TTS close behind. It also reports characters per second and a pronunciation-robustness score for names, codes, numbers and dates, which matters more than Elo for support lines and agents. Arena votes favour pleasant short clips; test long-form narration yourself.
04
Text to speech: time to first audio for agents
For voice agents, what counts is how quickly the first sound arrives after text is sent, not total generation time. Vendors quote their own figures (ElevenLabs gives about 150 ms to first speech for v4 Turbo), measured on their own networks; test from your region over a streaming WebSocket. Look for pinned model snapshots, as Cartesia offers, so a voice does not change under you.
05
Voice cloning and consent safeguards
Cloning a voice from a short sample is now common, so check what the provider does to prevent misuse: consent verification, watermarking (Google's SynthID, Resemble's Perth watermark in Chatterbox) and regional limits (Gemini's voice replication is unavailable in the EEA and UK). Open models usually rely on the user's conduct; Chatterbox has no consent check. If you clone real people, get written consent regardless of the tool.
06
Languages, and how well each is covered
A long language list is not the same as good quality in each language. Parakeet v3 covers 25 European languages and nothing else; Whisper accepts 99 but its card claims strong results in only about 10. Check the per-language results in the model card or test your target languages, especially for code-switching such as Hinglish.
07
Price units and open weights
Speech-to-text is priced per minute or hour of audio, text-to-speech per character or per audio token, which makes comparison awkward; Artificial Analysis normalises STT to dollars per 1,000 minutes and TTS to dollars per 1M characters. Watch for scheduled rises (Gemini 3.8 Flash TTS doubles in January 2027) and separate streaming rates. Open weights (Whisper MIT, Parakeet CC BY 4.0, Kokoro Apache 2.0, Chatterbox MIT) are cheapest at volume and keep audio on your servers, at some cost in accuracy or naturalness.
How the options differ
Speech models split by direction. Speech-to-text (STT) models turn audio into text and are judged on error rate, speed and features such as speaker labels; text-to-speech (TTS) models turn text into voice and are judged on naturalness, latency and voice control. Most vendors sell both, but rarely lead in both.
Within each, hosted APIs (ElevenLabs, Google, OpenAI, Deepgram, Cartesia) lead on quality and offer streaming for real-time agents. Open-weight models (Whisper, Parakeet, Kokoro, Chatterbox) run on your own hardware or cheap hosts, which suits privacy-sensitive or very high-volume work, but trail on accuracy and expressiveness and usually lack built-in streaming or safeguards.
11 tools, compared
Ranked by our rating. Open any review for sources, prices and the full verdict.
| Tool | Rating | Our verdict | Best for | Price |
|---|---|---|---|---|
| Eleven v4 and v4 Turbomodel | 4.7/5 | Nothing on the public leaderboards sounds better, but Google's Gemini 3.8 Flash TTS gets within 60 Elo points for about a fifth of v4's price. | Audiobook, game and dubbing producers who need the most expressive voices, and voice-agent teams who want top-ranked quality at Turbo's lower rate. | API: v4 $0.08 and v4 Turbo $0.04 per 1,000 characters. Subscriptions from Starter $6/month to Business $990/month include character allowances. |
| Gemini 3.8 Flash TTSmodel | 4.5/5 | Fifth in blind votes at under half Eleven v4 Turbo's price, Gemini 3.8 Flash TTS is a bargain with an expiry date of 1 January. | Developers producing long-form narration or multilingual dubbing who want near-top quality and can absorb January's price rise. | Flash TTS: $0.50 per 1M text tokens in, $9 per 1M audio tokens out (about $0.81 per hour of audio) until 31 Dec 2026, doubling on 1 Jan 2027. Flash-Lite: $6 per 1M audio tokens out. Batch half price; free tier in AI Studio. |
| Cartesia Sonic 3.6model | 4.4/5 | Sonic 3.6 closed most of the gap to ElevenLabs in one release and costs less per character than Eleven v4, but it trains on customer content unless you opt out. | Teams building production voice agents that want stable, versioned voices at a moderate price. | Free plan 20K credits/month (no commercial licence); Pro $5 (100K), Startup $49 (1.25M), Scale $299 (8M); overage $38 to $65 per 1M credits. One credit is one character, about 750 to 800 per minute of audio. |
| ElevenLabs Scribe v2model | 4.4/5 | Scribe v2 transcribes accurately and does the redaction and speaker work for you, but Microsoft now sells better accuracy for less than half the price. | Meeting, media and clinical products that need speaker labels and redaction as well as an accurate transcript. | API: Scribe v2 and v2 Medical $0.22 per hour of audio; Scribe v2 Realtime $0.39 per hour; keyterm prompting adds $0.05 per hour. Plans from $6/month include hours. |
| Gemini 3.5 Transcribemodel | 4.2/5 | Accurate and cheap enough for most transcription, yet no longer the obvious buy on either measure. | Teams already on the Gemini API who want file and streaming transcription from one vendor at about half a cent a minute. | Transcribe about $0.005 per minute ($2 input / $12 output per 1M tokens); Transcribe Live about $0.009 per minute ($3.50 / $21). Free tier in AI Studio and the API. |
| Parakeet TDT 0.6B v3model | 4.1/5 | Parakeet v3 is the speed record in transcription, with Whisper-class accuracy and a hard border around Europe. | Developers transcribing large volumes of European-language audio who prize speed and low cost. | Free under CC BY 4.0. Together AI's hosted version costs $0.0015 a minute ($1.50 per 1,000 minutes). |
| Deepgram Nova-3model | 4.0/5 | Nova-3 is quick and cheap enough for live call audio, but on a neutral test it makes more than twice the errors of the accuracy leaders. | Contact-centre and live-captioning teams that need fast, high-concurrency transcription of phone audio, including on their own servers. | Pre-recorded $0.0043/min (monolingual) or $0.0052/min (multilingual) pay-as-you-go; streaming $0.0048/min on promotion, $0.0077 regular. Growth plan from $4,000/year is cheaper. $200 free credit. |
| Kokoro-82Mmodel | 3.9/5 | Nothing this small sounds this good, but mid-table quality, preset voices and a model unchanged since January 2025 cap its ambitions. | Developers who need cheap, permissively licensed speech for narration, accessibility or prototypes. | Free under Apache 2.0. Hosted at $0.65 per million characters on Replicate (via Artificial Analysis), the cheapest on its leaderboard, and $0.99 on DeepInfra. |
| GPT Transcribemodel | 3.9/5 | A sensible upgrade for OpenAI shops, though the deprecation it rides on takes away speaker labels and timestamps that this model does not give back. | Teams on the OpenAI API who must leave Whisper or GPT-4o Transcribe before February 2027 and do not need speaker labels or timestamps. | $0.0045 per minute of audio ($4.50 per 1,000 minutes); no Batch support. Streaming gpt-live-transcribe is $0.017 per minute. |
| Whisper large-v3-turbomodel | 3.7/5 | Whisper turbo is now the cheap, unrestricted baseline that newer speech models are measured against, not the model that wins. | Teams that need cheap, permissively licensed transcription they can self-host or buy for cents an hour, and can live with second-tier accuracy. | Free under the MIT licence. Hosted versions are cheap: Artificial Analysis lists Groq at about $0.67 per 1,000 minutes. |
| Chatterboxmodel | 3.7/5 | Chatterbox is the open voice cloner a compliance team can sign off, thanks to its watermark, though not the one a listener would choose. | Teams that need commercially usable open voice cloning with provenance marking and can accept mid-table voice quality. | Free under the MIT licence. Artificial Analysis lists hosted Chatterbox at $25 and Resemble's Chatterbox HD at $40 per million characters. |
Side by side
The specifications that matter for this step. Checked 2026-10-19; confirm on the vendor’s site before you commit.
| Tool | Direction | Open weights | Languages | Streaming | AA score | Price |
|---|---|---|---|---|---|---|
| Eleven v4 and v4 Turbo | TTS | No | 90+ | Yes (v4 Turbo, ~150 ms) | 1328-1329 Elo | $80/1M chars; Turbo $40 |
| Gemini 3.8 Flash TTS | TTS | No | Not published | Near real-time (Flash-Lite) | 1278 Elo | About $16.50/1M chars (AA) |
| Cartesia Sonic 3.6 | TTS | No | 44 | Yes (built for agents) | 1287 Elo | About $49/1M chars |
| ElevenLabs Scribe v2 | STT | No | 90+ | Yes (Scribe v2 Realtime) | 2.2% AA-WER | $0.22/hour; Realtime $0.39 |
| Gemini 3.5 Transcribe | STT | No | 85+ | Yes (Transcribe Live) | 2.6% AA-WER | About $0.005/min |
| Parakeet TDT 0.6B v3 | STT | Yes, CC BY 4.0 | 25 European | Not published | 4.5% AA-WER | $1.50/1,000 min hosted |
| Deepgram Nova-3 | STT | No (enterprise self-host) | Multilingual | Yes | 5.2% AA-WER | $0.0043/min batch |
| Kokoro-82M | TTS | Yes, Apache 2.0 | 8 | Not published | 1064 Elo | $0.65/1M chars hosted |
| GPT Transcribe | STT | No | Not published | Yes (gpt-live-transcribe) | 3.3% AA-WER | $0.0045/min; live $0.017 |
| Whisper large-v3-turbo | STT | Yes, MIT | 99 | No (batch only) | 4.6% AA-WER | $0.67/1,000 min on Groq |
| Chatterbox | TTS | Yes, MIT | 23 | Turbo variant for agents | 1026 Elo (open model) | Free; HD hosted $40/1M chars |
Head to head and ranked guides
Where this stage fits
- Video production: we would start with Eleven v4 and v4 Turbo.
- Voice agent: we would start with Parakeet TDT 0.6B v3, Cartesia Sonic 3.6.