The default general-purpose TTS API. Eleven v3 covers 70+ languages with expressive audio-tag control but is too slow for realtime; Eleven v3 Conversational (~280 ms) and Flash v2.5 (~75 ms model-side, 32 languages) are the agent-grade options. Voice cloning is self-serve: instant cloning from about a minute of audio, professional cloning from ~30 minutes. Billing is in credits (1 credit per character on full-quality models, 0.5 on Flash/Turbo), which makes the effective rate tier-dependent; the published API rate is $0.10/1k chars for v3 and Multilingual v2, $0.05/1k for Flash, Turbo and v3 Conversational.
| Modality | TTS |
|---|
| Latency | ~75 ms TTFA (Flash v2.5); ~280 ms (v3 Conversational); v3 not realtime |
|---|
| Languages | 70+ (v3), 32 (Flash v2.5), 29 (Multilingual v2) |
|---|
| Voice cloning | Yes — instant (~1 min audio) and professional (~30 min), self-serve |
|---|
| Streaming | Yes — HTTP chunked and WebSocket |
|---|
| Unit price | $50–$100 per 1M characters |
|---|
Watch out: The flagship v3 model cannot be used for realtime conversation — you must fall back to Flash v2.5, which supports 32 languages instead of 70+ and sounds noticeably flatter. Credit-based billing means your effective per-character cost varies by subscription tier, making cost modelling harder than with flat per-minute vendors, and overage rates are steep. At scale, Cartesia, Rime and Inworld are 2–5x cheaper per million characters for comparable agent-grade quality. Concurrency is capped by tier, which bites hard on Starter/Creator.
Free 10k credits/mo; Starter $6/mo, Creator $11 first month then $22, Pro $99/mo (990k chars), Scale $299/mo, Business $990/mo. API: $0.10/1k chars (v3, Multilingual v2), $0.05/1k chars (Flash v2.5, Turbo, v3 Conversational). Agents/Speech Engine $0.08/min.
ElevenLabs' transcription line, sold separately from its TTS. Scribe v2 is batch-oriented, handles files up to 10 hours and 3 GB, diarizes up to 32 speakers, and accepts up to 1,000 keyterms for domain vocabulary. Scribe v2 Realtime streams at roughly 150 ms with a 50-keyterm limit. It currently sits near the top of independent WER leaderboards (~2.2% on Artificial Analysis) at a price close to commodity batch vendors, which is unusual — most accuracy leaders cost several times more.
| Modality | STT |
|---|
| Latency | ~150 ms (Scribe v2 Realtime); batch is offline |
|---|
| Languages | 90+ (accuracy varies widely by language) |
|---|
| Voice cloning | N/A |
|---|
| Streaming | Yes — Scribe v2 Realtime |
|---|
| Unit price | $0.22/hr batch, $0.39/hr realtime |
|---|
Watch out: Accuracy is very uneven across the claimed 90+ languages — English, Spanish and German land under 5% WER while Amharic and Zulu sit in the 25–50% range, so the headline language count oversells low-resource coverage. Realtime at $0.39/hr is roughly 1.3x AssemblyAI's streaming rate and 80x Deepgram Nova-3 streaming, so it is the wrong choice for high-volume voice agents. Credit-based billing on subscriptions obscures the true per-hour cost. HIPAA requires a sales conversation.
$0.22/hour (Scribe v2 batch), $0.39/hour (Scribe v2 Realtime). Add-ons: entity detection +$0.07/hr, keyterm prompting +$0.05/hr. Consumed as credits on subscription tiers (330 credits/min).
Eleven Music v2 generates full tracks with vocals and lyrics from a text prompt, exposed through the same API and credit system as ElevenLabs TTS rather than only through a consumer app. It bills 900 credits per minute of output, which works out to about $0.15 per minute at API rates. Unlike Suno and Udio it was built with label partnerships rather than after litigation, so commercial use is contractually clearer, and it is the only one of the three with a first-class programmatic API you can drop into a product pipeline.
| Modality | Music generation |
|---|
| Latency | Seconds to minutes per track; not realtime |
|---|
| Languages | Multilingual lyrics (English, Spanish, German, Japanese and others) |
|---|
| Voice cloning | No |
|---|
| Streaming | Yes — streaming music endpoint available |
|---|
| Unit price | $0.15 per minute of output |
|---|
Watch out: Musical quality and structural coherence lag Suno for full-song output — it is stronger at beds, loops and cues than at hit-shaped songs. There is no artist-voice cloning and no stem separation comparable to Suno Studio. At $0.15/min it is materially more expensive per finished minute than a $24/mo Suno Premier subscription if you are generating heavily. Documentation and model iteration cadence lag the TTS line.
$0.15 per minute of generated music via API (900 credits/min). Commercial rights begin at the Starter tier ($6/mo); free tier output is non-commercial.
Released July 2026 to replace the whisper-1 and gpt-4o-transcribe families, which OpenAI has scheduled for shutdown on 26 February 2027. gpt-transcribe handles completed files and batch workloads at $0.0045/min (about $0.27/hour) — 25% below the whisper-1 rate — and roughly halves whisper-1's word error rate on Common Voice across 22 languages. gpt-live-transcribe is the low-latency streaming variant at $0.017/min. Both accept context hints (keywords, language) to bias recognition on proper nouns, numbers and jargon.
| Modality | STT |
|---|
| Latency | Sub-second streaming (gpt-live-transcribe); batch is offline |
|---|
| Languages | 57 |
|---|
| Voice cloning | N/A |
|---|
| Streaming | Yes — gpt-live-transcribe |
|---|
| Unit price | $0.27/hr batch, $1.02/hr live |
|---|
Watch out: 57 languages is well below Whisper's 99 and Deepgram/Google's 80–100+, so it is a poor fit for long-tail language coverage. gpt-live-transcribe at $0.017/min is roughly 3.5x Deepgram Nova-3 streaming and 2.6x Deepgram Flux, which makes it expensive for always-on voice agents. Speaker diarization support is weaker than AssemblyAI's or Deepgram's. And you are migrating twice inside 18 months if you are still on whisper-1 — OpenAI has deprecated its own transcription lineup three times since 2025.
gpt-transcribe $0.0045/minute (~$0.27/hour); gpt-live-transcribe $0.017/minute (~$1.02/hour). Legacy whisper-1 and gpt-4o-transcribe $0.006/min, gpt-4o-mini-transcribe $0.003/min — all retiring 26 Feb 2027.
Two different products in one billing surface. The Realtime API (gpt-realtime, gpt-realtime-2.1, and the mini variants) is a single full-duplex speech-to-speech model — no separate STT or TTS stage — with built-in turn detection and tool calling, billed in audio tokens rather than minutes. The synthesis-only path is gpt-4o-mini-tts ($0.60/1M text input tokens, $12/1M audio output tokens, with steerable delivery instructions in the prompt) and the legacy tts-1 / tts-1-hd at $15 and $30 per 1M characters. Collapsing the pipeline removes two network hops and preserves prosody and interruption handling that a cascaded STT→LLM→TTS stack loses.
| Modality | Realtime speech-to-speech + TTS |
|---|
| Latency | Sub-second conversational turn-taking; ~300 ms for TTS-only calls |
|---|
| Languages | Inherits GPT multilingual coverage (~50+); no published per-language quality table |
|---|
| Voice cloning | No — fixed voice roster only |
|---|
| Streaming | Yes — WebRTC and WebSocket |
|---|
| Unit price | ~$0.06/min audio (gpt-realtime); $12–$30 per 1M chars for TTS-only |
|---|
Watch out: There is no voice cloning and no custom voices at all: you get a fixed voice roster, which rules it out for any brand-voice requirement. Audio-token billing is genuinely hard to forecast — long conversations re-send accumulated context on every turn, so real bills routinely run several times a naive per-minute estimate. You give up the ability to swap the LLM independently of the voice, and you lose the intermediate transcript unless you explicitly request it. tts-1 is an aging model and materially worse than Cartesia, Rime or ElevenLabs Flash for the same money.
gpt-realtime and gpt-realtime-2.1: $32/1M audio input tokens, $64/1M audio output tokens (roughly $0.06/min of audio at list, before re-sent text context). gpt-realtime-mini / 2.1-mini: $10/$20 per 1M. gpt-4o-mini-tts: $0.60/1M text in, $12/1M audio out. tts-1 $15/1M chars, tts-1-hd $30/1M chars.
Still the default open-weight ASR baseline three years on, and the only widely deployed model with genuine 99-language coverage. large-v3 sits around 7.4% average WER on the Open ASR Leaderboard at 1.55B parameters; large-v3-turbo trades a little multilingual accuracy for roughly 8x faster decoding and fits in about 6 GB of VRAM. The MIT licence permits commercial use and redistribution with no restrictions, and the ecosystem (faster-whisper, whisper.cpp, WhisperX) provides quantisation, word timestamps and diarization that upstream lacks.
| Modality | STT |
|---|
| Latency | Batch only; large-v3-turbo ~8x realtime on a modern GPU |
|---|
| Languages | 99 |
|---|
| Voice cloning | N/A |
|---|
| Streaming | No native streaming; community wrappers only |
|---|
| Unit price | $0 (self-hosted compute only) |
|---|
Watch out: No native streaming — every realtime Whisper deployment is a community chunking wrapper that reintroduces latency and boundary errors, so it is the wrong choice for voice agents. It hallucinates fluent text on silence and non-speech audio, a well-documented failure mode that has caused real problems in clinical settings. No built-in diarization, no word-level timestamps without WhisperX, and no punctuation quality on par with 2026 commercial models. Newer open models (Parakeet, Voxtral, Canary) now beat it on English accuracy at a fraction of the parameter count.
Free (MIT weights). Self-hosted GPU cost only — typically well under $0.05/hour of audio on a rented L4/A10 at reasonable batch sizes. OpenAI's hosted whisper-1 endpoint is $0.006/min but retires 26 Feb 2027.
· open source
NVIDIA's NeMo ASR family, and the reason self-hosted transcription got cheap. parakeet-tdt-0.6b-v3 covers 25 European languages, beats Whisper large-v3 on measured accuracy at roughly a quarter of the parameters, and processes audio around 6.5x faster than comparable-accuracy open models — throughput is the headline, with batch RTFx figures in the thousands on a single datacentre GPU. Weights are CC-BY-4.0, and NeMo ships cache-aware streaming variants so unlike Whisper it can genuinely do low-latency streaming without a chunking hack.
| Modality | STT |
|---|
| Latency | Cache-aware streaming variants available; batch throughput 1000x+ realtime on datacentre GPUs |
|---|
| Languages | 25 (European) for TDT v3 |
|---|
| Voice cloning | N/A |
|---|
| Streaming | Yes — NeMo cache-aware streaming models |
|---|
| Unit price | $0 (self-hosted compute only) |
|---|
Watch out: 25 European languages only — it is not a substitute for Whisper if you need Arabic, Hindi, Vietnamese or long-tail languages. Deployment is meaningfully harder than Whisper: NeMo has heavier dependencies, a smaller community ecosystem, and far less third-party tooling for diarization and timestamps. CC-BY-4.0 imposes an attribution requirement that MIT does not. It is strongly optimised for NVIDIA hardware, so CPU or Apple Silicon inference is a poor experience compared with whisper.cpp.
Free (CC-BY-4.0 weights). Self-hosted GPU cost only; throughput is high enough that per-hour compute cost is typically a fraction of a cent at scale.
· open source
The price floor for hosted ASR. Nova-3 handles 80+ languages at $0.0043/min pre-recorded and $0.0048/min streaming — roughly a tenth of what most competitors charge — with a medical variant available. Flux is the newer piece and the more interesting one: an STT model with end-of-turn detection built into the model rather than bolted on as a VAD heuristic, aimed squarely at voice-agent pipelines, covering 10 languages at $0.0065/min. Aura-2 provides TTS at $0.030 per 1,000 characters. Deepgram also sells self-hosted and VPC deployments, which most competitors reserve for large enterprise contracts.
| Modality | STT + TTS |
|---|
| Latency | Flux: model-integrated end-of-turn detection for agents; Aura-2 sub-200 ms TTFB |
|---|
| Languages | Nova-3 80+; Flux 10; Aura-2 English and Spanish |
|---|
| Voice cloning | No — fixed Aura voice library |
|---|
| Streaming | Yes — WebSocket for both STT and TTS |
|---|
| Unit price | $0.26/hr STT batch; $0.29–$0.39/hr STT streaming; $30 per 1M chars TTS |
|---|
Watch out: Aura-2 has no voice cloning and a limited voice library confined mainly to English and Spanish, so it is not competitive with ElevenLabs or Cartesia for branded or multilingual synthesis. Nova-3's accuracy leads on clean conversational English but slips against Scribe v2 and Microsoft's MAI models on accented and noisy audio in independent benchmarks. Flux covers only 10 languages. Several of the headline streaming rates are promotional and revert higher (Nova-3 streaming lists at $0.0077/min at standard rates), so model your costs against the non-promotional column.
Nova-3 monolingual: $0.0043/min batch, $0.0048/min streaming ($0.26–$0.29/hr). Nova-3 multilingual: $0.0052/$0.0058 per min. Flux English streaming: $0.0065/min. Aura-2 TTS: $0.030/1k chars. Growth plan discounts ~10–20%; $200 free credit to start.
Positions itself one layer above raw transcription: alongside Universal-3.5 Pro ($0.21/hr async) and Universal-2 ($0.15/hr async) it sells speaker identification, translation, entity detection, sentiment, topic detection and custom formatting as separately priced per-hour add-ons that stack additively. Streaming runs from $0.15/hr for Universal-Streaming up to $0.45/hr for Universal-3.5 Pro Realtime. The add-on model is the differentiator — you can get IAB topic classification or 50-type entity detection without running a second LLM pass yourself.
| Modality | STT + speech understanding |
|---|
| Latency | ~300 ms streaming (Universal-Streaming immutable transcripts) |
|---|
| Languages | 99+ async; multilingual streaming model available |
|---|
| Voice cloning | N/A |
|---|
| Streaming | Yes — WebSocket, billed by session duration |
|---|
| Unit price | $0.15–$0.21/hr async, $0.15–$0.45/hr streaming, before add-ons |
|---|
Watch out: Streaming is billed by WebSocket session duration, not audio duration: a socket held open for 60 minutes carrying 30 minutes of speech bills 60 minutes, which can double your real cost if your client does not close idle connections. Add-ons stack additively and a fully loaded configuration can exceed $1/hr, well above the headline rate. No TTS at all, so it cannot be your single voice vendor. Legacy `best` and `nano` model identifiers are deprecated and silently reroute, which can change your output format without warning.
Async: Universal-3.5 Pro $0.21/hr, Universal-2 $0.15/hr. Streaming: Universal-Streaming (English or multilingual) $0.15/hr, Universal-3.5 Pro Realtime $0.45/hr. Add-ons per hour: diarization +$0.02 (async) / +$0.12 (streaming), translation +$0.06, entity detection +$0.08, topic detection +$0.15, medical mode +$0.15, voice focus +$0.10.
A state-space-model TTS house built for the voice-agent latency budget. Sonic 3.6 is generally available with 44 languages, holds the top Quality Elo (1282) on Artificial Analysis's text-to-speech arena, and varies pacing and intonation from transcript context without SSML tags. Cartesia bills 1 credit per character, so effective cost runs roughly $37/1M characters on the $299 Scale plan down to enterprise rates well below that. It also sells Ink STT and a bundled voice-agent product at $0.06/min. Dated model snapshots (e.g. sonic-3.6-2026-08-27) let you pin behaviour, which most TTS vendors do not offer.
| Modality | TTS (+ Ink STT) |
|---|
| Latency | ~90 ms model-side TTFA (Sonic 3); ~166–190 ms measured end-to-end |
|---|
| Languages | 44 |
|---|
| Voice cloning | Yes — instant (Pro tier) and professional (Startup tier) |
|---|
| Streaming | Yes — WebSocket with raw PCM output |
|---|
| Unit price | ~$37–$50 per 1M characters depending on tier |
|---|
Watch out: The sub-100 ms figures are model-side only — independent measurements put real-world TTFA around 166–190 ms once network transit is included, so do not budget on 40 ms. Credit accounting is convoluted: professional cloning costs a million credits up front and then permanently bills 1.5 credits per character, which can quietly inflate spend by 50%. Effective per-character cost on lower tiers (~$50/1M chars) is worse than ElevenLabs Flash and far worse than Inworld. Concurrency is only 2–3 requests on Free and Pro, so you cannot load-test meaningfully without paying up. Ink STT is less mature than the TTS line.
Free $0 (20k credits), Pro $5/mo (100k credits, ~133 TTS min), Startup $49/mo (1.25M credits), Scale $299/mo (8M credits ≈ $37/1M chars). 1 credit per character; Pro Voice Cloning costs 1M credits one-time and raises the rate to 1.5 credits/char. Voice agents $0.06/min plus $0.014/min telephony.
A UK ASR vendor with an unusually strong accented-English and dialect story, sold on a flat $0.129/hour Pro rate with no per-feature add-on pricing. Three transcription models are offered: Enhanced (highest accuracy), Standard (cheaper) and Melia 1 (multilingual). The distinguishing commercial features are genuine on-premises and air-gapped deployment available below enterprise contract size, automatic volume discounts above 500 hours/month per model, further discounts from 24,000 hours/year, and a 33% discount if you opt in to model training on your data. An English-only low-latency TTS is available with more languages promised.
| Modality | STT (+ English-only TTS) |
|---|
| Latency | Configurable realtime max-delay; sub-second achievable |
|---|
| Languages | 55+ languages and dialects |
|---|
| Voice cloning | No (custom voice development at enterprise tier only) |
|---|
| Streaming | Yes — realtime sessions, 50 concurrent on Pro |
|---|
| Unit price | $0.129/hr |
|---|
Watch out: At $0.129/hour it is roughly 5x Deepgram Nova-3 batch and 10x Google dynamic batch, so it loses badly on pure cost for commodity transcription. 55+ languages is mid-pack — below Deepgram, Google and Whisper. The TTS product is English-only and immature; treat Speechmatics as an STT vendor. The 33% training discount requires giving up data isolation, which contradicts the main reason most buyers choose them. Realtime concurrency is capped at 50 sessions on Pro, requiring an enterprise contract for anything larger.
Pro $0.129/hour base, 50 concurrent realtime sessions, 10 file jobs/sec; 20% discount available. Free tier: $100 credit, 2 concurrent realtime sessions. Automatic volume discounts above 500 hrs/month per model; 33% off for training opt-in. Enterprise custom.
A TTS-only vendor optimised almost exclusively for voice-agent turn latency and for sounding like real conversational speech rather than an audiobook narrator. Two models: Mist v3 at 37 ms P50 time-to-first-audio across 94 voices and 4 languages for $0.03/1k chars, and Coda at 96 ms P50 across 184 voices and 8 languages for $0.05/1k chars with word-level timestamps. Practical features that matter in contact-centre work are first-class: explicit character-by-character spelling control, continuous speed adjustment, and self-hosted deployment maintaining sub-100 ms latency. Enterprise adds on-prem/VPC, HIPAA BAA and SOC 2.
| Modality | TTS |
|---|
| Latency | 37 ms P50 TTFA (Mist v3); 96 ms P50 (Coda) |
|---|
| Languages | 4 (Mist v3), 8 (Coda); 600+ voices total |
|---|
| Voice cloning | Enterprise only — no self-serve cloning |
|---|
| Streaming | Yes — HTTP and WebSocket |
|---|
| Unit price | $30–$50 per 1M characters |
|---|
Watch out: Language coverage is the weakest of any major TTS vendor here — 4 languages on the fast model and 8 on Coda, versus 44 for Cartesia and 70+ for ElevenLabs — so it is simply unusable for most international products. Voice cloning is enterprise-only, with no self-serve path. There is no STT and no music, so it can never be a single-vendor solution. Support on the Starter plan is a public Slack channel. Voices skew heavily toward American English conversational registers, which is a poor fit for narration or media work.
Starter $0.03/1k chars (Mist v3, ~$0.03/min) or $0.05/1k chars (Coda), ~800 free minutes to start, 20 concurrent generations, no card required. Enterprise: custom volume pricing, unlimited concurrency and voice clones.
The only vendor here whose core thesis is prosody and emotional expression rather than latency or price. Octave 2 is an LLM-based TTS you steer with natural-language acting directions rather than SSML, running at roughly 100–200 ms model-side latency across about 11 languages. EVI (Empathic Voice Interface) is the speech-to-speech agent layer, which measures vocal expression in the user's speech and conditions the response on it — genuinely different behaviour from a standard STT→LLM→TTS cascade. Voice cloning is unlimited on all paid tiers, and voices can be created by text description rather than reference audio.
| Modality | TTS + speech-to-speech agent (EVI) |
|---|
| Latency | ~100–200 ms model-side (Octave 2) |
|---|
| Languages | ~11 |
|---|
| Voice cloning | Yes — unlimited on paid tiers, plus text-prompted voice design |
|---|
| Streaming | Yes |
|---|
| Unit price | ~$60–$70 per 1M characters; EVI ~$0.058/min at Pro |
|---|
Watch out: About 11 languages is far behind every other commercial TTS vendor on this list — English, Spanish, French, German, Italian, Portuguese, Russian, Arabic, Hindi, Japanese, Korean and little else. Latency is fine but not competitive with Rime or Cartesia for hard realtime budgets. Emotional steering is inconsistent run-to-run, which makes deterministic production output harder; expect to regenerate. Character allowances on lower tiers exhaust quickly and overage rates are not clearly published. The company is small relative to ElevenLabs or Deepgram, which is a real procurement consideration.
Free $0 (10k chars, 5 EVI min); Starter $3/mo (30k chars, 40 EVI min); Creator $7/mo (140k chars, 200 EVI min); Pro $70/mo (1M chars ≈ $0.07/1k, 1,200 EVI min ≈ $0.058/min); Scale $200/mo (3.3M chars ≈ $0.06/1k, 5,000 EVI min); Business $500/mo (10M chars, 12,500 EVI min); Enterprise custom.
Originated as a game-character AI company and repositioned into voice infrastructure, and now one of the strongest quality-per-dollar options on the Artificial Analysis TTS arena. TTS-2 lists at $25/1M characters and TTS-2 Flash at $15/1M on the on-demand tier, with plan-based discounts stepping down to $12.50 and $7 at the $1,500/mo Growth tier and as low as $5/1M at enterprise — roughly a third to a tenth of ElevenLabs' per-character rate for comparable agent-grade output. STT is offered at $0.10–$0.15/hour. It picked up substantial migration traffic after PlayHT's shutdown.
| Modality | TTS (+ STT) |
|---|
| Latency | Vendor claims ~100–200 ms TTFA; limited independent verification |
|---|
| Languages | ~12 |
|---|
| Voice cloning | Yes — custom voices, count capped by tier (100 to 30,000+) |
|---|
| Streaming | Yes |
|---|
| Unit price | $5–$25 per 1M characters; STT $0.10–$0.15/hr |
|---|
Watch out: The best rates require large monthly commitments ($1,500/mo Growth or an enterprise contract); at the on-demand rate it is only moderately cheaper than ElevenLabs Flash. Language coverage (roughly 12 languages) is well behind Cartesia and ElevenLabs. The company pivoted from game NPCs to voice infrastructure relatively recently, so the TTS line has a shorter production track record than Deepgram or ElevenLabs. Custom voice counts are gated by tier. Independent latency figures are thinner than for Rime or Cartesia — treat the vendor's sub-200 ms claims as unverified.
On-Demand (free to start, ~70 TTS min): TTS-2 $25/1M chars, TTS-2 Flash $15/1M. Creator $25/mo: $20 and $10. Builder $100/mo: $17.50 and $9. Developer $300/mo: $15 and $8. Growth $1,500/mo: $12.50 and $7. Enterprise as low as $5/1M. STT $0.15/hr on entry plans, $0.10/hr on Creator and above.
The hyperscaler option, and the cheapest credible batch STT on the market: Speech-to-Text V2 dynamic batch runs about $0.003/minute (~$0.18/hour) with a 24-hour turnaround, against $0.016/minute for standard realtime recognition falling to $0.004/minute above 2M minutes per month. On the synthesis side, Chirp 3 HD is $30/1M characters, Standard voices $4/1M, and Studio voices $160/1M, with Gemini-TTS available for prompt-steered delivery. The Chirp model family is included in V2 at no premium. The real argument for Google is not the models — it is VPC-SC, CMEK, data residency and one invoice.
| Modality | STT + TTS |
|---|
| Latency | Streaming recognition and synthesis supported; no published TTFA figures |
|---|
| Languages | STT 100+; TTS 50+ languages and locales |
|---|
| Voice cloning | Instant Custom Voice — allowlisted and consent-gated, not self-serve |
|---|
| Streaming | Yes — both STT and TTS |
|---|
| Unit price | TTS $4–$160 per 1M chars; STT $0.18–$0.96/hr |
|---|
Watch out: Voice quality and expressiveness trail ElevenLabs, Cartesia and Hume — Chirp 3 HD is competent but not the voice you pick when the voice is the product. Self-serve voice cloning does not exist: Instant Custom Voice is allowlisted and consent-gated. Pricing is genuinely hard to model, with per-voice-type tiers, separate V1/V2 API rates, and features that quietly double the bill. Streaming realtime STT at $0.016/min is over 3x Deepgram. Google has repeatedly renamed and re-versioned this product line (V1, V2, Chirp, Chirp 2, Chirp 3, Gemini-TTS), and migrations have not always been smooth.
TTS: Standard $4/1M chars, Chirp 3 HD $30/1M chars, Studio $160/1M chars; Chirp 3 HD includes a 1M char/month free tier. STT V2: dynamic batch ~$0.003/min (~$0.18/hr), standard realtime $0.016/min falling to $0.004/min at 2M+ min/month.
The enterprise incumbent, and the widest voice catalogue of any vendor here — hundreds of neural voices across 140+ locales. The pricing structure rewards batch heavily: $0.18/hour batch versus $0.36/hour fast transcription and $1.00/hour standard realtime, with commitment tiers reaching $0.50/hour at 50,000 hours/month. Microsoft Foundry now also hosts OpenAI's gpt-transcribe and gpt-live-transcribe alongside its own MAI-Transcribe models, which independently rank at or near the top of WER leaderboards (MAI-Transcribe-2 at ~2.0% AA-WER). Custom Neural Voice provides high-fidelity brand voice cloning but requires Microsoft approval and documented consent.
| Modality | STT + TTS |
|---|
| Latency | Realtime streaming recognition and synthesis; no published TTFA figure |
|---|
| Languages | STT 100+; TTS 140+ locales and 500+ voices |
|---|
| Voice cloning | Yes — Custom Neural Voice, gated behind Microsoft approval |
|---|
| Streaming | Yes — both directions |
|---|
| Unit price | STT $0.18–$1.00/hr; TTS ~$16 per 1M chars |
|---|
Watch out: Realtime STT at $1.00/hour is 20x Deepgram Nova-3 streaming and 3.5x Google — it is close to the most expensive realtime ASR on this list, and the $0.18/hour batch headline is what people quote while actually paying the realtime rate. Enhanced features are charged separately per hour per feature on realtime, which compounds fast. Custom Neural Voice requires an application, consent documentation and Microsoft approval, adding weeks to a project. The SDK and portal are noticeably heavier to work with than a single REST call to Deepgram or ElevenLabs. Product naming across Azure AI Speech, Microsoft Foundry and the MAI model line is currently confusing.
STT: batch $0.18/hr, fast transcription $0.36/hr, realtime $1.00/hr; custom models $1.20/hr realtime, $0.225/hr batch, plus $0.0538/model/hour endpoint hosting and $10/compute-hour training. Enhanced add-ons (diarization, language ID, pronunciation assessment) +$0.30/hr each for realtime, free on batch. Neural TTS ~$16/1M chars with 500k chars/month free. Free F0 tier: 5 audio hours/month.
The strongest permissively licensed open TTS in 2026 and the pragmatic answer when you cannot send audio to a vendor or cannot afford per-character pricing. Roughly 0.5B parameters under MIT, zero-shot voice cloning from 5–20 seconds of reference audio with no fine-tuning, an emotion-exaggeration parameter that ranges from monotone to theatrical, and ~200 ms synthesis latency on a GPU. Chatterbox Multilingual covers 23+ languages; Chatterbox Turbo adds faster inference and paralinguistic tagging for laughter and other non-speech sounds. Resemble also offers hosted inference if you would rather not run it, and independent blind tests have had Turbo beating ElevenLabs in preference tests.
| Modality | TTS |
|---|
| Latency | ~200 ms on GPU (Turbo faster) |
|---|
| Languages | 23+ (Multilingual variant) |
|---|
| Voice cloning | Yes — zero-shot from 5–20 seconds, no fine-tuning |
|---|
| Streaming | Yes — chunked streaming via community and Resemble tooling |
|---|
| Unit price | $0 (self-hosted compute only) |
|---|
Watch out: MIT licensing on the weights does not remove your liability for cloning a voice without consent — the model will happily clone anyone from 5 seconds, and the legal and ethical burden shifts entirely to you. Quality is competitive but not consistently ahead of the best commercial models, and it is noticeably less stable on long-form text, unusual punctuation and numbers. You inherit all the ops: GPU capacity planning, cold starts, streaming chunking, and pronunciation dictionaries that commercial vendors provide. 23 languages is behind Cartesia and ElevenLabs. There is no SLA, no support contract and no roadmap commitment for the open weights.
Free (MIT weights, no royalties or usage caps). Self-hosted GPU cost only. Managed hosting available via Resemble AI at separate paid rates.
· open source
Formerly one of the four or five vendors any TTS shortlist would include, with a large voice library and self-serve cloning. Meta acquired PlayHT/PlayAI in July 2025 and folded the roughly 35-person team into Superintelligence Labs. The API went dark on 26 July 2025, weeks ahead of the announced schedule, and the full platform sunset on 31 December 2025 — user accounts, saved audio and voice clones were deleted at the cutoff. The play.ht domain no longer resolves. The technology is now used internally at Meta for AI Characters and wearables and is not available as a standalone service. Listed here because it still appears in comparison articles and older tutorials.
| Modality | TTS (discontinued) |
|---|
| Latency | N/A — service shut down |
|---|
| Languages | N/A — service shut down |
|---|
| Voice cloning | N/A — existing clones were deleted 31 Dec 2025 |
|---|
| Streaming | N/A — API offline since July 2025 |
|---|
| Unit price | N/A — no longer sold |
|---|
Watch out: The product does not exist. There is no API, no dashboard and no data recovery path — voice clones and generated audio were deleted at the 31 December 2025 cutoff. If you have inherited a codebase calling the PlayHT API, migrate to Cartesia, Rime, Inworld or ElevenLabs; several of those vendors published explicit PlayHT migration guides during the shutdown. Treat this as the standard cautionary example of building on a single-vendor voice API without an abstraction layer.
unknown — no longer sold; the service is shut down
· discontinued
The market leader in generative music by a wide margin, producing full tracks with vocals, lyrics and structure from a text prompt. The current lineup runs v4 through v5.5, with Premier adding Suno Studio and advanced stem separation. Pro is $8/month for 2,500 credits (about 500 songs) with commercial use rights and 20 downloads/month; Premier is $24/month for 10,000 credits and 60 downloads. Suno settled Warner Music Group's copyright suit in November 2025 and will launch licensed models during 2026, at which point Suno has stated the current models will be deprecated.
| Modality | Music generation |
|---|
| Latency | Roughly 30–60 seconds per track; not realtime |
|---|
| Languages | Multilingual lyrics; English strongest |
|---|
| Voice cloning | No — personas from your own uploads only, no artist voice cloning |
|---|
| Streaming | Playback streams while generating in-app; no streaming API |
|---|
| Unit price | $8/mo for ~500 songs; $24/mo for ~2,000 songs |
|---|
Watch out: The models you build on today are explicitly slated for deprecation: Suno has said current models will be retired when the Warner-licensed models ship in 2026, so output style and quality will change under you. Universal Music and Sony litigation is still live with no trial date, so the copyright position is unresolved for two of the three majors. Downloads are capped even on paid plans (20/month on Pro, 60/month on Premier) despite generous credit allowances — the credit count is not the real limit. Credits expire monthly and do not roll over. There is no official public generation API, so programmatic integration means a subscription and unofficial tooling. Free-tier output carries no commercial rights.
Free $0 (50 daily credits, ~10 songs/month, v4.5-all only, no commercial rights); Pro $8/mo (2,500 credits ≈ 500 songs, 20 downloads, commercial rights, v5.5); Premier $24/mo (10,000 credits ≈ 2,000 songs, 60 downloads, Suno Studio). Credits do not roll over.
Suno's closest quality rival before the litigation, particularly strong on vocal timbre and genre fidelity. Udio settled with Universal Music Group on 29 October 2025 and with Warner Music Group subsequently, and is rebuilding as a fully licensed platform with UMG and WMG catalogues, restricted to artists who opt in. The cost of that settlement was severe: Udio disabled all downloads on 30 October 2025, opened a 48-hour recovery window on 3–5 November 2025, and has not restored export since. The relaunched joint platform is described as a walled garden where generated audio cannot be downloaded or posted elsewhere. As of late August 2026 neither company had confirmed a public launch date.
| Modality | Music generation |
|---|
| Latency | Seconds to minutes per track; not realtime |
|---|
| Languages | Multilingual lyrics |
|---|
| Voice cloning | No |
|---|
| Streaming | In-app playback only; no export since Oct 2025 |
|---|
| Unit price | unknown |
|---|
Watch out: You cannot get your audio out. Downloads have been disabled since October 2025 and the announced relaunch is explicitly a walled garden with no export or off-platform posting, which makes Udio unusable for any commercial pipeline that needs a deliverable file. The relaunch has slipped through 2026 with no confirmed date. Sony's claims remain active. Current models are to be replaced by licensed versions, so anything you produce now may not be reproducible later. Confidence in every figure here is low precisely because the product is mid-restructure — verify directly before committing.
unknown — subscription tiers exist but the product is mid-transition and the pricing page could not be verified; do not plan against current published rates
· beta