- tts
- stt
- offline
- open-source
- supertonic
- kokoro
- grok
- gemini
- realtime-api
- on-device
- whisper
- voice-cloning
- browser
- arena
If 2025 was the year TTS became “usable,” the first half of 2026 is when it hit an inflection point. On-device engines are approaching cloud quality faster than most predicted. Cloud APIs are collapsing the three-stage STT→LLM→TTS pipeline into a single call. Open-source models are covering capabilities — emotion control, multi-speaker dialogue, zero-shot voice cloning — that were exclusive to commercial APIs just months ago. And all of this is happening simultaneously.
Here is the TTS/STT landscape as of mid-2026, based on offlinetts.com’s blog coverage.
1. On-Device TTS: Two Flagship Releases and a Steady Workhorse #
1.1 Supertonic — 99M Parameters, 31 Languages, 167× Real-Time
Supertonic (by Supertone Inc.) is the most technically impressive on-device TTS release of the season. It runs on ONNX Runtime for fully local inference.
| Dimension | Spec |
|---|---|
| Model size | ~99 million parameters (99M) |
| Inference speed | 167× real-time — 1 hour of audio in ~22 seconds |
| Languages | 31 |
| Run targets | Raspberry Pi, phone, browser (WebGPU / WASM) |
| Latest version | Supertonic 3: improved accuracy and language coverage |
| Custom voices | Voice Builder tool from short reference clips |
OfflineTTS integrated Supertonic on May 13, initially supporting 5 languages (English, Spanish, Portuguese, French, Korean) with 10 preset voice styles. Two core knobs control the quality/speed trade-off — Steps (denoising passes) and Speed (speech rate):
| Steps | Best For | Trade-off |
|---|---|---|
| 4–8 | Quick drafts and previews | Fastest, less refined |
| 8 (default) | Everyday use | Balanced quality and latency |
| 12–20 | Final exports | Smoother but slower |
Model assets are cached in IndexedDB after first load, so repeat visits skip re-downloads.
Bottom line: For any use case where privacy, latency, or offline operation is a requirement, Supertonic has become the benchmark that other on-device solutions are measured against.
1.2 ToBe SAID — Fully Offline Android TTS Engine
Unlike Supertonic’s developer focus, ToBe SAID delivers a polished end-user experience as a system-level Android TTS engine. It runs entirely on-device and is built around ebook-to-audiobook conversion. Recent updates significantly improved voice stability and naturalness. The free tier covers one voice slot; Pro unlocks unlimited slots.
1.3 Kokoro TTS — The On-Device Workhorse
Kokoro is built on StyleTTS 2 architecture — 82M parameters, 82MB ONNX model, and the most mature ecosystem among browser-runnable TTS engines:
88 voices, graded A–D by quality** 9 languages**: English (US/UK), Japanese, Chinese, French, Spanish, Hindi, Italian, Portuguese, Korean** Run modes**: browser (WebGPU / WASM), Python CLI, API ($0.65/1M chars)** License**: Apache 2.0
Top English A-grade voices:
| Voice | Accent | Best For |
|---|---|---|
| Heart | American | Audiobook narration, YouTube voiceovers |
| Bella | American | Conversational content, podcasts |
| Nova | American | Professional presentations, e-learning |
| Nicole | British | British narration, formal content |
TTS Arena performance:
Kokoro ranks 32nd overall (Elo 1056) out of 74 models — but it is #1 among browser-runnable models. It excels at knowledge-sharing content (articles, docs, education) with Elo 1066, outperforming Google WaveNet (873) and Amazon Polly Neural (868).
| Category | Accent | Elo | Win Rate |
|---|---|---|---|
| Knowledge sharing | All | 1065.8 | 57.1% |
| Assistants | All | 1065.8 | 50.9% |
| Customer service | US | 1135.4 | 46.0% |
For anyone who wants TTS without cloud APIs or GPU servers, Kokoro is the most pragmatic choice available.
2. Cloud APIs: Consolidation, Speed, Emotion #
xAI Grok TTS & STT — A New Contender
xAI entered the voice API market in May with Grok TTS and STT, emphasizing speed, multilingual accuracy, and easy integration (already available through Telnyx for telephony apps). The message is clear: major AI platform companies now treat voice I/O as a core capability, not an add-on.
OpenAI Realtime API — Single-Call Architecture
This is the most architecturally significant shift this year. The traditional voice agent pipeline:
STT (speech→text) → LLM (reasoning+generation) → TTS (text→speech)
Each hop adds latency and integration overhead. The Realtime API compresses this into one call: speech in, speech out.
| Dimension | Traditional Pipeline | OpenAI Realtime API |
|---|---|---|
| API calls | 3+ | 1 |
| Latency | Cumulative per hop | End-to-end optimized |
| Orchestration | High (WebSocket management) | Low |
| Best fit | When cloud is acceptable | Privacy-sensitive still needs on-device |
Gemini 3.1 Flash TTS — Emotional Control at Scale
Google’s Gemini 3.1 Flash TTS brings two breakthroughs:
70+ languages— one of the broadest cloud TTS coverages available** Fine-grained emotional control**— adjusts tone, emphasis, and affect to match content context
This upgrades TTS from “listenable” to “engaging.” In the Artificial Analysis TTS Arena it ranks #2 with 1205.8 Elo, just 4 points behind Inworld Realtime TTS 1.5 Max (1209.6).
Other Commercial API Notes
ElevenLabs v3: Quality benchmark (1178 Elo), but $100/1M chars — most expensive** MiniMax Speech 2.8 HD / Turbo**: Solid top-10 performers (1163.7 / 1146.7 Elo)** Azure Speech Service**: Expanding language coverage, enterprise SLAs
3. TTS Arena Leaderboard Deep Dive #
The Artificial Analysis Speech Arena ranks 74 models by Elo through blind A/B listening tests — the most objective TTS quality benchmark available.
Top 10 (All Closed-Source Commercial)
| Rank | Model | Creator | Elo | Win Rate | Price/1M chars |
|---|---|---|---|---|---|
| 1 | Realtime TTS 1.5 Max | Inworld | 1209.6 | 73.3% | $35 |
| 2 | Gemini 3.1 Flash TTS | 1205.8 | 72.4% | $36.61 | |
| 3 | Eleven v3 | ElevenLabs | 1178.0 | 68.9% | $100 |
| 4 | Inworld TTS 1 Max | Inworld | 1165.4 | 66.1% | $35 |
| 5 | Speech 2.8 HD | MiniMax | 1163.7 | 65.2% | $100 |
| 6 | Realtime TTS 1.5 Mini | Inworld | 1158.4 | 66.2% | $25 |
| 7 | Step TTS 2 | StepFun | 1149.1 | 64.6% | $40 |
| 8 | Speech 2.8 Turbo | MiniMax | 1146.7 | 64.0% | $60 |
| 9 | Speech 2.6 HD | MiniMax | 1133.5 | 62.1% | $100 |
| 10 | Speech 2.6 Turbo | MiniMax | 1128.7 | 61.3% | $60 |
Open-Weight Model Standings
| Overall Rank | Model | Creator | Elo | Win Rate | Notes |
|---|---|---|---|---|---|
| 11 | Fish Audio S2 Pro | Fish Audio | 1128.7 | 61.0% | Best open-weight; needs GPU |
| 32 | Kokoro 82M v1.0 | Kokoro | 1056.2 | 54.4% | Browser-runnable, 82MB |
| 33 | Voxtral TTS | Mistral | 1055.9 | 52.3% | Released March 2026 |
| 35 | Maya1 | Maya Research | 1050.6 | 50.5% | |
| 51 | Fish Speech 1.5 | Fish Audio | 1011.9 | 49.1% | |
| 52 | Chatterbox | Resemble AI | 1006.4 | 47.9% | |
| 55 | Zonos v0.1 | Zyphra | 1000.0 | 47.1% | |
| 60 | OpenVoice v2 | OpenVoice | 949.9 | 44.0% | |
| 66 | XTTS v2 | Coqui | 885.9 | 36.4% | |
| 67 | StyleTTS 2 | StyleTTS | 878.8 | 37.4% | |
| 74 | MetaVoice v1 | MetaVoice | 765.2 | 21.5% |
The Gap Is Shrinking — Fast
| Era | Best Open-Weight Elo | Best Commercial Elo | Gap |
|---|---|---|---|
| 2023 | 879 (StyleTTS 2) | 1102 (OpenAI TTS-1) | 223 |
| 2024 | 950 (OpenVoice v2) | 1107 (ElevenLabs v2) | 157 |
| Early 2025 | 1006 (Chatterbox) | 1134 (MiniMax Speech 2.6 HD) | 128 |
| Mid 2025 | 1056 (Kokoro) | 1170 (Eleven v3 pre) | 114 |
| Early 2026 | 1129 (Fish Audio S2 Pro) | 1210 (Inworld RT 1.5 Max) | 81 |
From 223 to 81 — a 64% reduction in under 3 years. At the current trajectory, an open-weight model could crack the top 10 within a year.
⚠️
Caveat:Elo measures listener preference, not objective quality. It does not capture latency, privacy, cost at scale, language coverage, or browser-deployability — factors that matter enormously in production.
4. Open-Source Model Ecosystem: MOSS-TTS and Beyond #
MOSS-TTS Family (OpenMOSS / MOSI.AI)
The most ambitious open-source speech generation effort today — not a single model but a family:
| Variant | Core Capabilities |
|---|---|
| MOSS-TTS | High-fidelity long-form speech, multi-speaker dialogue, real-time streaming TTS, sound design |
| MOSS-TTS-Nano | Lightweight variant for mobile and embedded devices |
Key capabilities:
Long-form fidelity— sustained quality across extended passages, solving the “starts good, degrades after a few sentences” problem** Multi-speaker dialogue**— distinct voices within a single generation, no stitching needed** Real-time streaming**— chunk-by-chunk output with minimal buffering** Sound effects**— non-speech audio generation for games and media production
Also active: Chatterbox, Fish Audio S2 Pro, Kokoro, and Zonos all continue to receive community updates.
5. Browser-Based STT: The Technical Landscape #
Browser-based speech recognition has moved from demo-grade to production-ready in 2026.
How It Works
Three stages:
Audio capture & decoding— microphone or file → mono 16kHz PCM (Whisper’s expected format), WebCodecs for hardware acceleration** Neural inference**— Whisper encoder-decoder via ONNX Runtime Web, WebGPU preferred, WASM fallback** Post-processing**— raw tokens → text, timestamp tokens → segment boundaries, long audio via 30s sliding windows
Model Family Comparison
| Library | Models | Languages | Model Size | WebGPU | Streaming | Worker-Based |
|---|---|---|---|---|---|---|
| transformers.js | tiny/base/small/large | 99 | 40–3000 MB | ✅ | ❌ | ❌ |
| browser-whisper | tiny/base/small | 99 | 40–240 MB | ✅ | ✅ | ✅ |
| Whisper.cpp | tiny~large-v3 | 99 | 39–3000 MB | ❌ | ❌ | Native |
| Moonshine | tiny/base | English only | 6–61 MB | ✅ | ❌ | ❌ |
| Distil-Whisper | small/medium | English only | 185–760 MB | ✅ | ❌ | ❌ |
Model Size vs. Accuracy
| Model | Parameters | Download (hybrid quant) | Relative Accuracy | WebGPU RT Factor |
|---|---|---|---|---|
| Whisper Tiny | 39M | ~40 MB | Adequate for clear speech | 10–15× |
| Whisper Base | 74M | ~76 MB | Best balance for most use cases | 5–8× |
| Whisper Small | 244M | ~240 MB | Handles accents/noise best | 2–4× |
Why Quantization Matters More Than You Think
ONNX models can be quantized to reduce size, but not all parts should be quantized equally:
Encoder (feature extractor): Quantization-sensitive —** fp32 recommended**. Quantizing to q8 degrades feature quality, producing garbled output on accented or noisy audio.** Decoder (text generator)**: Quantization-tolerant — q4 or q8 both work; q4 is significantly smaller.
This is why browser-whisper defaults to hybrid quantization (fp32 encoder + q4 decoder). A full q8 model at ~300 MB is not just larger — it can produce worse transcriptions than the 76 MB hybrid version, because encoder quantization noise propagates through the entire decoder stack.
WebGPU vs WebAssembly
WebGPU provides 5–10× speedup over WASM for Whisper, but browser support is uneven:
| Browser | WebGPU Support |
|---|---|
| Chrome 113+ / Edge 113+ | ✅ Best performance |
| Safari 17.4+ | ✅ macOS and iOS |
| Firefox | ❌ Behind a flag |
| Linux Chrome | ⚠️ Needs --enable-unsafe-webgpu |
A robust production implementation must fall back to WASM gracefully.
Two Common Issues with Long Audio
Hallucination: Whisper can generate repetitive nonsense text at chunk boundaries or in silent regions. In transformers.js v3.8.1,SuppressTokensLogitsProcessor
is commented out, leaving 90 hallucination-prone tokens unsuppressed. browser-whisper applies correct pipeline configuration to mitigate this. - Timestamp drift: At chunk boundaries, timestamps can drift or overlap, requiring post-processing for subtitle formats (SRT/VTT).
Production Architecture Recommendations
Most use cases: browser-whisper + whisper-base + hybrid quantization** Maximum accuracy**: browser-whisper + whisper-small** Slow connections / fastest first load**: whisper-tiny** Real-time streaming (live captions): Moonshine Server-side deployment**: Whisper.cpp
6. Voice Cloning: Three Engines Compared #
Voice cloning is now fully achievable offline. OfflineTTS compared three leading engines in detail.
At a Glance
| Dimension | KokoClone | Kitten Fine-Tune | Piper Express Clone |
|---|---|---|---|
| Method | Zero-shot (speaker encoder) | Full fine-tuning | Synthetic data + fine-tune |
| Reference audio | 3–10 seconds | 5–30 min (transcribed) | 3–10 seconds |
| Training needed? | ❌ No | ✅ 6–12 hours | ✅ 2–4 hours |
| GPU needed? | ❌ CPU works | ✅ 8–40GB VRAM | ✅ Recommended 8–12GB VRAM |
| Clone quality | Good | Good (with enough data) | Best |
| Inference speed | ~150ms/10s text (CPU) | Very fast (tiny model) | Real-time on CPU |
| Model size | ~84MB | 20–30MB | ~75MB |
| Multi-language | ✅ 9 languages | ❌ Single | ✅ |
Detailed Breakdown
KokoClone — Zero-Shot Instant Cloning
KokoClone uses an ECAPA-TDNN speaker encoder to extract acoustic features from a short audio sample, producing a speaker embedding that plugs directly into Kokoro’s existing decoder. No retraining needed.
from kokoclone import KokoClone
clone = KokoClone(device="cpu")
audio = clone.text_to_speech(
text="Hello, this is my cloned voice.",
ref_wav="my_voice.wav",
language="en"
)
Best for: Rapid prototyping, personal assistants, IoT devices. Clone quality depends heavily on reference audio quality.
Kitten TTS — Fine-Tuning for Tiny Footprints
Kitten is a lightweight VITS architecture (15M params, 15–80MB). No zero-shot mechanism — you need 5–30 minutes of transcribed audio and 6–12 hours of GPU training. The output is a 20–30MB model that runs extremely fast on CPU.
Best for: Embedded systems, mobile apps, Raspberry Pi Zero — where size and speed matter more than setup convenience.
Piper Training Suite — Express Clone
Two-stage pipeline:
Chatterbox generates 1,500+ synthetic training clips from 3–10 seconds of reference audioPiper fine-tuning runs 300–500 epochs, exports a standard ONNX model
python cloneToPiper.py MyVoice ./reference.wav \
--samples 200 --epochs 500 --quality high --language en-us
piper -m ./exports/MyVoice.onnx -t "This is my cloned voice"
Highest quality among the three, but requires 2–4 hours of training.
Best for: Audiobook production, customer service voice bots, game character voiceovers.
Decision Framework
**Need the voice now?→ KokoClone (seconds, no GPU) Highest possible quality?→ Piper Express Clone (fine-tuning > zero-shot) CPU only?→ KokoClone (only option) Embedded / smallest footprint?**→ Kitten (20–30MB)
7. Developer Decision Guide #
Combined across all the trends above, here’s how to choose your stack in mid-2026:
| Your Scenario | TTS Pick | STT Pick | Rationale |
|---|---|---|---|
| Offline / privacy non-negotiable | Supertonic / Kokoro | Whisper variant (browser-whisper) | Quality is production-grade; data never leaves the device |
| Highest quality, cloud OK | Gemini 3.1 Flash / ElevenLabs v3 | OpenAI Whisper API | Best emotional expressiveness or overall naturalness |
| Building a voice agent, fast | OpenAI Realtime API (single call) | Built-in | Eliminates 3-service orchestration |
| Browser-first user experience | Kokoro / Supertonic (both WebGPU) | browser-whisper | Zero install, zero config, fully offline |
| Low-cost at scale | Kokoro self-hosted ($0.65/1M chars) | Whisper.cpp | 100×+ cost reduction vs commercial APIs |
| Voice cloning (quick) | Kokoro + KokoClone | — | 3–10 seconds of reference audio |
| Voice cloning (high quality) | Piper Training Suite | — | Train once, deploy everywhere |
| Embedded / IoT | Kitten TTS / MOSS-TTS-Nano | Moonshine | Tiny models, extremely fast inference |
8. Summary #
The most important trend of H1 2026 is not any single release — it’s convergence:
On-device quality is catching up to cloud.The Elo gap has shrunk from 223 to 81 — a 64% reduction — and open-weight models could crack the top 10 within a year.
Cloud APIs are simplifying architectures.The three-stage pipeline is becoming a single call, making voice agent development dramatically easier.
Open-source models now cover formerly commercial-only capabilities.Emotion control, multi-speaker dialogue, zero-shot voice cloning — all achievable locally.
The browser is becoming a viable runtime.WebGPU enables 82MB TTS and 76MB STT models to run in a browser tab in real-time, offline, with zero installation.
The result is a TTS/STT ecosystem that offers better options at every point on the quality–cost–privacy spectrum. Whether you are building a voice assistant, an audiobook platform, a video dubbing tool, game characters, or an accessibility feature:
If you haven’t put voice in your product yet, 2026 is the year to start.
Based on offlinetts.com blog coverage (April–May 2026). Sources include official announcements from xAI, OpenAI, Google, and Supertone; the Artificial Analysis Speech Arena leaderboard; GitHub repository activity; and community discussions.
Reference Articles (from offlinetts.com)
| Title | Date |
|---|---|