Versus
ElevenLabs vs OpenAI TTS
Flagship commercial TTS head-to-head: quality, cost, latency, voice library.
Read the comparison →Codesota · Speech · Vol. IIThe register of speech-to-text and text-to-speechIssue: April 22, 2026
§ 00 · Speech
Speech AI, both directions.
Two pillars share this register. Speech-to-text now clears human accuracy on clean audio; text-to-speech clears the blind-test bar for naturalness. We keep both on the same page because the pipeline almost always needs both.
18 STT models and 18 TTS models tracked, sourced from the shared model catalogue. Shaded rows mark current state of the art. Numbers shown only where reported; every model links to paper or code where available.
§ 01 · Speech-to-text
Word error rate, ranked.
LibriSpeech test-clean remains the canonical benchmark. Lower is better. Human-annotator WER on this split sits in the 2–4% band, which several models now clear.
- Metric
- WER · lower is better
- Models
- 18 tracked · top 8 shown
- Dataset
- LibriSpeech test-clean
Top 8 · April 2026
Shaded row marks current SOTA
| # | Model | Vendor | Kind | Params | Trend | WER | Δ |
|---|---|---|---|---|---|---|---|
| 01 | Parakeet RNNT 1.1B | NVIDIA | Open Source | 1.1B | 1.8 | — | |
| 02 | Conformer XL | Research | 600M | 2.0 | +0.2 | ||
| 03 | Deepgram Nova-3 | Deepgram | Cloud API | — | 2.2 | +0.2 | |
| 04 | Voxtral Large | Mistral AI | Cloud API | — | 2.3 | +0.1 | |
| 05 | AssemblyAI Universal-2 | AssemblyAI | Cloud API | — | 2.4 | +0.1 | |
| 06 | Canary 1B | NVIDIA | Open Source | 1B | 2.4 | 0.0 | |
| 07 | Whisper Large v3 Turbo | OpenAI | Open Source | 809M | 2.5 | +0.1 | |
| 08 | Gladia v2 | Gladia | Cloud API | — | 2.5 | 0.0 |
Fig 1 · WER on LibriSpeech test-clean. Δ is difference against the row above. Sparkline shows a directional trendline from each model's vendor-reported trajectory — not a per-submission history.
§ 02 · Text-to-speech
Mean opinion score, ranked.
Naturalness is scored by human raters on a 1–5 scale. Commercial and open-source entries now overlap in the 4.5–4.8 band — a gap small enough that the right model is chosen by latency, licence or voice cloning rather than raw quality.
- Metric
- MOS · higher is better
- Models
- 18 tracked · top 8 shown
- Evaluation
- Subjective listening tests
Top 8 · April 2026
Shaded row marks current SOTA
| # | Model | Vendor | Kind | Params | Trend | MOS | Δ |
|---|---|---|---|---|---|---|---|
| 01 | ElevenLabs Turbo v2.5 | ElevenLabs | Cloud API | — | 4.8 | — | |
| 02 | Sesame CSM | Sesame | Open Source | 1B+ | 4.7 | -0.1 | |
| 03 | OpenAI TTS HD | OpenAI | Cloud API | — | 4.7 | 0.0 | |
| 04 | Gemini 2.5 Pro TTS | Cloud API | — | 4.7 | 0.0 | ||
| 05 | Cartesia Sonic 2 | Cartesia | Cloud API | — | 4.7 | 0.0 | |
| 06 | ElevenLabs Flash v2.5 | ElevenLabs | Cloud API | — | 4.6 | -0.1 | |
| 07 | PlayHT 3.0 | PlayHT | Cloud API | — | 4.6 | 0.0 | |
| 08 | Orpheus TTS | Canopy Labs | Open Source | 3B | 4.6 | 0.0 |
Fig 2 · MOS is subjective. Vendors publish different listener panels and reference tracks; direct comparison below 0.1 should be treated as noise.
§ 03 · Comparison pages
Pairwise, and by use-case.
Long-form reads for the common decisions: which commercial TTS, which open-source, which model fits podcasts, audiobooks, voice bots or cloning.
Versus
ElevenLabs vs Cartesia Sonic
Quality leader against the purpose-built low-latency challenger.
Read the comparison →Versus
OpenAI TTS vs Google Cloud TTS
Hyperscaler comparison — pricing, voices, SSML, streaming.
Read the comparison →Use-case
Best TTS for podcasts
Long-form naturalness ranked: pacing, breath, intonation over 30+ minutes.
Read the comparison →Use-case
Best TTS for audiobooks
SSML, character voices, consistency across chapters.
Read the comparison →Use-case
Best TTS for real-time voice agents
TTFB under 200ms: Cartesia, ElevenLabs Flash, Gemini Flash.
Read the comparison →Use-case
Best TTS for voice cloning
Zero-shot similarity, data requirements, and consent-ethics framing.
Read the comparison →Open
Best open-source TTS
Kokoro, Sesame CSM, Orpheus, F5-TTS, Dia — licensed and deployable.
Read the comparison →Fig 3 · Each comparison page has its own evidence table; these are editorial reads, not benchmark duplicates.
§ 04 · Featured deep-dive
How speech becomes a picture.
Eleven open-source TTS voices, the same prompt, rendered through five DSP lenses and Griffin-Lim resynthesis. A reproducible walkthrough of the representations that vocoders, ASR systems and human ears actually read — mel spectrograms, MFCC, F0, formants.
Every figure is generated from the same code path; every voice is labelled with its provenance. No fabricated spectrograms, no stock audio. If the sample cannot be reproduced, it doesn't appear.
§ 05 · Benchmarks
The datasets we believe.
Canonical for each direction plus the community-adopted follow-ups. LibriSpeech, Common Voice and VCTK are canonicalised in our dataset registry; FLEURS, AudioBench and EARS are tracked qualitatively pending canonicalisation.
Rows with a mark live in the registry and carry full lineage.
| Benchmark | Scope | Primary metric | Year | Source | |
|---|---|---|---|---|---|
| LibriSpeech | Speech-to-Text | wer-test-clean | 2015 | link → | |
| Common Voice | Speech-to-Text | wer | 2019 | link → | |
| LJ Speech | Text-to-Speech | mos | 2017 | link → | |
| VCTK | Text-to-Speech | mos | 2019 | link → | |
| TTS Intelligibility | Text-to-Speech | critical-entity-accuracy | 2026 | link → | |
| FLEURS | Speech-to-Text | WER (per-lang) | 2022 | link → | |
| AudioBench | Audio-LLM | composite | 2024 | link → | |
| EARS | Text-to-Speech | MOS · subjective | 2024 | link → |
Fig 5 · Solid marker = canonicalised in the Codesota registry. Hollow marker = widely cited, tracked qualitatively, not yet graded.
ASR · English
1.8WER, ↓
TTS · naturalness
4.8MOS, ↑
Realtime TTS
~90ms TTFB, ↓
Open-source TTS
4.7MOS, ↑
Fig 6 · Directional trends across four speech axes. Dot marks the current SOTA entry from the catalogue.
§ 06
How it works
Two pipelines, one register.
Modern speech recognition takes raw audio into mel-spectrogram features, runs them through a Conformer or Transformer encoder, and decodes with CTC, RNNT or attention. Post-processing — language-model rescoring, punctuation, diarisation — yields the final transcript.
Modern speech synthesis runs the pipeline in reverse. Text is embedded by a language model; acoustic tokens are predicted autoregressively or by flow matching; a vocoder or neural codec decodes those tokens back to waveform. The neural audio codec — EnCodec, SoundStream, Mimi — is the hinge that lets TTS borrow the tooling of LLMs.
What changed recently is the representation. Once audio could be tokenised, every architectural trick from text generation became available to speech: pretraining, instruction-tuning, prompted style control, zero-shot cloning. That is why the open-source gap in TTS closed so quickly after 2023.
On the STT side, the Conformer block — self-attention plus convolution — is still the workhorse. Whisper took a different path with a pure Transformer encoder-decoder trained on weak supervision at scale, trading some efficiency for massive multilingual coverage.
Related
Neighbouring registers.
Other modality hubs on Codesota worth reading next.