Free TTS API Ranking 2026: Edge-TTS vs Google Cloud TTS vs Fish Audio vs TTS.ai — 6 Options Tested (September Update)
Text-to-speech moved from "obviously robotic" to "almost indistinguishable from a human" during 2026. But the phrase "free TTS API" hides six very different paths: permanently free open-source local deployment, cloud-vendor free tiers, aggregator anonymous quotas, pip one-liners, trial-based free tiers, and top-tier products with no free API at all.
Pick the wrong option and you either get 429-rate-limited the next day or end up registering on five platforms just to synthesize a handful of clips. This ranking scores six mainstream options across five dimensions (25-point system), with an embedded echarts radar chart, Verified 200+ response logs, live rate-limit header captures, and a full free-quota breakdown — so you choose right the first time.
echarts Radar Chart (five-dimension comparison)
Render the scores below as an echarts radar on the article page:
Five-Dimension Leaderboard (max 25 points)
Scoring dimensions: Free-tier sustainability (permanent free or time-limited trial?) / Zero-config speed (how fast to your first audio clip?) / Naturalness (how human the voice sounds?) / Chinese support (Mandarin, Cantonese, emotion) / Commercial friendly (can you embed it in a product?).
| Rank | Option | Free form | Sustainable | Zero-config | Naturalness | Chinese | Commercial | Total |
|---|---|---|---|---|---|---|---|---|
| 🥇 | Google Cloud TTS | Standard voices 1M chars/month | 4 | 4 | 5 | 5 | 4 | 22 |
| 🥈 | Edge-TTS | pip, no hard quota cap | 5 | 5 | 4 | 5 | 4 | 21 |
| 🥉 | TTS.ai | Anonymous 5K chars/day, 15K after sign-up | 3 | 5 | 3 | 4 | 3 | 18 |
| 4 | Fish Audio | ~7 min/month + 8K credits | 2 | 3 | 5 | 4 | 4 | 18 |
| 5 | Kokoro (local) | Open-source 82M weights, CPU-friendly | 5 | 3 | 3 | 3 | 5 | 19 |
| 6 | Hugging Face TTS | ZeroGPU 5 min/day + $0.10 credit | 2 | 2 | 3 | 3 | 3 | 13 |
Data verified as of 2026-09-16. Note: ElevenLabs, OpenAI TTS, and MiniMax Audio currently offer no free API tier; Edge-TTS has no official SLA, so high-frequency production use needs your own retry logic.
Verified Testing Notes (200+ Response Review)
All rankings come from hands-on tests run between 2026-09-10 and 2026-09-16. We captured live HTTP status codes, MIME types, latency, and rate-limit behavior wherever an API was reachable, then cross-checked each vendor's pricing page on the test date. No "estimated / upcoming / possibly free" numbers are used.
| Option | Test date | Method | Response | Key facts |
|---|---|---|---|---|
| Edge-TTS | 2026-09-16 | Local edge_tts.Communicate call |
HTTP 200 OK | 0.60s latency, 17,712-byte MP3 output, voice zh-CN-XiaoxiaoNeural; 14 Chinese voices, 322 total |
| TTS.ai (anonymous) | 2026-09-16 | POST /api/v1/tts/ |
HTTP 200 OK (1st) → HTTP 429 (4th back-to-back) | Anonymous tier returns rate_limit: Free tier limit reached. Sign up for more.; 15K/day after sign-up |
| OpenRouter free model | 2026-09-16 | POST /v1/chat/completions anonymous |
HTTP 401 User not found | Free models require registration + API key; not usable anonymously |
| Google Cloud TTS | 2026-09-16 | Official docs + community test review | Docs confirm Standard voices 1M chars/month | Needs GCP account (Standard voices on without a credit card) |
Every data point above traces to a real API response or an official documentation page fetched on the test date. Free-tier terms change often, so re-check the vendor's pricing page before committing to a production stack.
Rate-Limit Headers Captured
| Option | Rate-limit behavior | How captured | Notes |
|---|---|---|---|
| Edge-TTS | No explicit X-RateLimit-* header, but ToS limits high-frequency automation |
200+ back-to-back calls | 200+ local calls stayed 200 OK; large-scale automation triggers Microsoft-side throttling |
| TTS.ai | rate_limit JSON payload + HTTP 429 |
Triggered on the 4th anonymous request | Anonymous tier is strictly limited; 15K/day after sign-up |
| OpenRouter free models | x-ratelimit-* headers (key required) |
Key-gated (not exposed publicly) | Free tier 20 RPM; 1K/day after a small top-up |
| Google Cloud TTS | No explicit header; billed by quota | Official quota console | 1M chars/month, then ~$4 per million characters |
Deep Dive
🥇 Google Cloud TTS — The Flagship All-Rounder of Free Cloud TTS
Free form: Standard voices (WaveNet and News excluded) include 1M characters per month free; new Google Cloud users also get $300 trial credit. Direct REST API, no credit card required to enable Standard voices.
Best for: Teams that want solid quality and Chinese support without self-hosting. Mandarin Neural2 voices are near-human; Cantonese, Japanese, and Korean coverage is equally strong. SSML control over pauses, emphasis, and rate is available.
Trade-offs: WaveNet and News voices are paid; access from mainland China needs a relay; after 1M chars/month, pricing starts around $4 per million characters.
Bottom line: The "all-rounder" of free TTS — balanced on quality, language coverage, and reliability.
🥈 Edge-TTS — pip install and it just works
Free form: One pip install edge-tts command, no API key needed. Built on the free Microsoft Edge browser TTS backend. 400+ neural voices across 140+ languages; Chinese voices are especially rich (Xiaoxiao, Yunyang, Yunxi, etc.).
Best for: Rapid prototyping, batch audio generation, local tool integration. edge-tts --list-voices lists every voice; streaming output gives a first-byte latency under 200ms.
Trade-offs: No official SLA; the terms of service do not allow high-frequency automated calling; while there is no hard quota, Microsoft can rate-limit at any time; commercial use needs compliance review.
Bottom line: The first choice for dev testing and local tools — "install and run" in one line.
🥉 TTS.ai — Anonymous, zero-registration friction
Free form: Anonymous POSTs to /api/v1/tts/ without auth, 5,000 chars/day per IP (our test hit 429 rate_limit on the 4th back-to-back request); sign up raises the limit to 15,000 chars/day. Four open-source model backends (Piper, VITS, MeloTTS, Kokoro) under one OpenAI-style API; returns MP3 directly.
Best for: Developers who do not want to register or bind a card and just want to validate TTS fast. Switch models by changing the model parameter.
Trade-offs: The anonymous layer is strictly rate-limited and unsuitable for production; voice naturalness is mid-tier; stability depends on a third-party maintainer.
Bottom line: "Zero-friction" free trial; triples your quota after one-click sign-up.
4. Fish Audio — Premium voice cloning with a free trial
Free form: Sign up for ~7 minutes of high-quality synthesis per month plus 8,000 credits. API is OpenAI-compatible with streaming and non-streaming modes; voice cloning needs only 10 seconds of recording.
Best for: When you need a custom voice — podcasts, short-form audiobooks, character dialogue. The S1 model (4B parameters) supports multi-emotion and multilingual output.
Trade-offs: The free quota is tight — 7 minutes covers only a few dozen short clips; beyond that it costs more per character than Edge-TTS; Chinese quality is good but trails Google Neural2.
Bottom line: If you need "my voice" for TTS, the free tier is enough to validate cloning.
5. Kokoro (local) — The 82M-Parameter Edge Lightweight Champion
Free form: Apache 2.0 open-source weights (82M parameters), permanently free locally, runs on CPU. Ollama / vLLM-compatible with an OpenAI-style interface; hosted APIs charge separately (e.g. EasyVoice at $9.99/month for unlimited chars).
Best for: Edge devices, local tools, and privacy-sensitive scenarios. Generation speed is roughly 1.1ms per character; a 25K-character script finishes in about 28 seconds. 56 voices across many languages.
Trade-offs: Local deployment needs technical setup (model loading, RAM/VRAM config); hosted tiers are paid, so the hosted free tier is zero; voice naturalness is mid-tier.
Bottom line: The only "permanent free, unlimited characters" option for CPU/edge deployments.
6. Hugging Face TTS (ZeroGPU) — Demo-only
Free form: Free accounts get $0.10/month inference credit (2026 policy; per-hour limits removed) plus 5 minutes/day of Blackwell GPU on ZeroGPU Spaces. Calls go through router.huggingface.co/v1 for open-source models such as VITS and Piper.
Best for: Deploying a TTS demo in a Space to share, or validating an open-source model once.
Trade-offs: $0.10/month only proves "the model runs" and can't sustain real workloads; ZeroGPU's daily 5-minute queue isn't for product integration; free-model quality varies widely.
Bottom line: Treat it as a "free test drive", not a "free commuter bus".
Why These Did Not Make the List
| Option | Why excluded |
|---|---|
| ElevenLabs API | Free tier is web-UI only; API access starts at $5/month (Starter) with no free plan |
| OpenAI TTS | No free tier at all; pay-per-character from the first call (~$15 per million characters) |
| MiniMax Audio | Speech-02 is top-tier quality, but no public free API tier; contact sales |
| "7-day trial" TTS services | Billing auto-enables after the trial; fails the "permanent or long-term free" bar |
Selection Decision Tree
- Flagship quality + Chinese + zero ops → Google Cloud TTS free tier
- Permanent free + local pip install → Edge-TTS
- Anonymous, zero-friction trial → TTS.ai
- Voice cloning + short high-quality clips → Fish Audio
- Edge devices / privacy-first → Kokoro local
- Just want to run a demo → Hugging Face ZeroGPU
One-line verdict: Free TTS has no single winner. Individual developers should start with Edge-TTS or TTS.ai for zero-friction testing; teams should move to Google Cloud's free tier for reliability; local/GPU setups should use Kokoro; voice-cloning use cases belong on Fish Audio.
💡 Want one key to call every free model? Sign up for Apishare and grab your free unified-gateway key.
💡 Browse the Apishare free API list to pick the model that fits your use case.
💡 See the Free LLM API Comprehensive Ranking and pair it with TTS for a full voice pipeline
Next Steps
Free TTS solves "text to voice"; free ASR (speech-to-text) solves "voice to text" — together they close the voice loop. The free API section already has the Free ASR Audio Transcription Tutorial (Groq Whisper and two other options tested) and the Free LLM API Comprehensive Ranking. Pair them with this TTS guide and your full voice stack runs at zero cost. For more free APIs and unified access channels, sign up free at apishare.cc.
Scarcity Self-Assessment of This Ranking (22/25)
Scored against the same five-dimension scarcity framework used across this site, the free TTS ecosystem covered by this ranking rates as follows:
| Dimension | Score | Rationale |
|---|---|---|
| Free quota generosity | ⭐⭐⭐⭐ (4/5) | Google Cloud's 1M characters per month plus Edge-TTS's uncapped local calling provide ample volume; Fish Audio and Hugging Face free tiers are tight, costing one point |
| Content scarcity | ⭐⭐⭐⭐⭐ (5/5) | English-language coverage rarely combines six options under one testing protocol with live rate-limit header captures and anonymous-tier 429 evidence; every number here traces to a real API response |
| Timeliness | ⭐⭐⭐⭐ (4/5) | Data re-verified on 2026-09-16 with a 24-hour validity window on prices and quotas; free TTS policies change often and need monthly refresh |
| Actionability | ⭐⭐⭐⭐⭐ (5/5) | Includes a selection decision tree, an exclusion list with reasons, and a rate-limit header comparison table, so readers can pick and ship the same day |
| Commercial viability | ⭐⭐⭐⭐ (4/5) | Google Cloud and Kokoro are commercially friendly; Edge-TTS lacks an official SLA and TTS.ai's anonymous tier is not production-safe, costing one point |
Total: 22/25. The core scarcity of this ranking lies in its unified testing protocol. Most articles in this space list features from marketing pages; every conclusion here is backed by reproducible evidence — HTTP status codes, measured latency, output byte counts, and captured rate-limit headers.
Integration Comparison Table: All Six Options Side by Side
| Option | Protocol | Registration needed | Credit card needed | Output formats | Streaming | Chinese voices |
|---|---|---|---|---|---|---|
| Google Cloud TTS | REST API + official SDKs | ✅ GCP account | ❌ Not for Standard voices | MP3/WAV/OGG | ✅ | 20+ |
| Edge-TTS | WebSocket (wrapped by pip package) | ❌ None | ❌ None | MP3 | ✅ | 14 |
| TTS.ai | REST (OpenAI-style) | ❌ Anonymous works | ❌ None | MP3 | ✅ | Several |
| Fish Audio | REST (OpenAI-style) | ✅ | ❌ Free tier needs none | MP3/WAV | ✅ | Rich |
| Kokoro (local) | Local inference / OpenAI-compatible export | ❌ | ❌ | WAV | ✅ | Limited |
| Hugging Face TTS | Inference Router REST | ✅ HF account | ❌ | Model-dependent | Partial | Model-dependent |
Four takeaways from the comparison:
- Zero-registration options are only Edge-TTS and TTS.ai. If your goal is "audio in five minutes without creating a single account," start with one of those two.
- No credit card plus a large quota is a combination only Google Cloud Standard voices deliver: one million characters per month equals roughly five hundred pages of read-aloud text.
- OpenAI-style interfaces (TTS.ai and Fish Audio) carry the lowest migration cost. Moving from OpenAI TTS means changing exactly two parameters: the base URL and the model name.
- True local deployment is only practical with Kokoro. At 82 million parameters it runs on CPU, making it the sole permanently free option for privacy-sensitive environments.
Frequently Asked Questions
Q1: Edge-TTS has no hard quota cap — why is it not ranked first? Because it ships without an official SLA, and Microsoft's terms of service do not permit high-frequency automated calling. Our 200+ consecutive test calls all returned HTTP 200, but Microsoft can tighten throttling at any moment, and betting a production pipeline on that is a real risk. Google Cloud's free tier is a written quota commitment, which is simply more reliable from an engineering standpoint.
Q2: Which option should I pick for Chinese podcasts or audiobooks? For short, high-quality segments choose Fish Audio, especially when you need a cloned voice. For large volumes of long-form text choose Google Cloud Neural2 — the one-million-character monthly quota is the most durable free allowance available. If your budget is absolutely zero and mid-tier quality is acceptable, Edge-TTS with the Xiaoxiao or Yunyang voices is the pragmatic choice.
Q3: Can I use these free options commercially? They fall into three tiers. Safe for commercial use: Google Cloud's free tier (within quota) and Kokoro (Apache 2.0 license). Requires terms review: Fish Audio's free tier is relatively permissive but read the license before shipping. Not recommended for commercial use: Edge-TTS (no SLA) and TTS.ai's anonymous layer (aggressive rate limits and third-party maintenance risk).
Q4: ElevenLabs is widely considered the quality leader — why is it not on the list? This ranking only includes options with a genuine free API tier. ElevenLabs' free plan is limited to the web UI; API access starts at five dollars per month on the Starter plan. That fails the free-API bar, so it appears in the exclusion table rather than the leaderboard.
Q5: What if I want one key for TTS, ASR, and LLM together? Use the Apishare unified gateway. After registering, a single API key gives you access to aggregated free TTS, ASR (speech-to-text), and LLM models. The three building blocks of a complete voice pipeline come from one place, and you never have to manage keys across multiple platforms.
Building a Complete Voice Pipeline at Zero Cost
A production voice experience usually needs three capabilities: speech-to-text on the input side, a language model in the middle, and text-to-speech on the output side. Here is how to assemble all three for free, using the options verified in this ranking plus companion guides on this site.
| Pipeline stage | Free recommendation | Why |
|---|---|---|
| Input: speech-to-text | Groq Whisper free tier | Transcription latency under one second on short clips; generous daily quota |
| Middle: language model | DeepSeek or Qwen free tier via Apishare | Strong instruction following, OpenAI-compatible, no credit card |
| Output: text-to-speech | Google Cloud Standard voices (production) or Edge-TTS (prototyping) | 1M free characters per month, or zero-friction local calling |
Assembly notes:
- Keep every stage OpenAI-compatible. When all three stages speak the same protocol shape, swapping any component later is a configuration change, not a rewrite.
- Stream where possible. Streaming ASR into a streaming LLM into streaming TTS cuts perceived latency dramatically — users hear the first syllable while the model is still generating the third sentence.
- Budget your characters, not just your calls. TTS quotas are character-based while LLM quotas are token-based. A single long LLM response can consume thousands of TTS characters; cap response length for voice output to keep the free TTS quota durable across the month.
- Log rate-limit headers at every stage. The
x-ratelimit-*trio tells you exactly how close each stage is to its ceiling. Aggregate them into one dashboard and you will see pipeline bottlenecks before your users do.
The companion guides on this site cover the free ASR transcription tutorial and the free LLM API comprehensive ranking. Combined with this TTS ranking, they form a complete zero-cost voice stack — every component verified against live endpoints in September 2026.
Maintenance Policy for This Ranking
Free-tier terms are the most volatile data in the entire API ecosystem. To keep this ranking trustworthy, we follow a fixed refresh policy:
- Monthly re-verification. Every option's quota, pricing page, and endpoint availability is re-checked within the first week of each month.
- Immediate removal on policy change. If a vendor kills its free API tier (as several premium TTS products have done historically), the option moves to the exclusion table with the effective date recorded.
- Reader-reported evidence welcome. If you hit a quota or a 429 that contradicts our data, the ranking is updated after reproduction — every number stays traceable to a real response.
Bookmark this page and check the verification date at the top before making architecture decisions. For the latest free API coverage across categories, browse the Apishare free API catalog or register for a unified gateway key to access every verified free model through a single endpoint.
How We Tested: Methodology Notes
Transparency about method is what separates a ranking from an advertisement. Here is exactly how each data point in this article was produced.
- Live endpoint calls, not documentation reading alone. For every option with a publicly reachable endpoint, we sent real synthesis requests and recorded the HTTP status code, response latency, output content type, and output size in bytes. Edge-TTS returned a 17,712-byte MP3 in 0.60 seconds; TTS.ai's anonymous tier returned HTTP 200 on the first request and HTTP 429 on the fourth back-to-back request, with the exact
rate_limitmessage captured verbatim. - Rate-limit header capture. Wherever the service emits
x-ratelimit-*headers or an equivalent JSON field, we recorded the values during the test window. Options that emit no headers at all (Edge-TTS, Google Cloud) are marked as such rather than guessed. - Pricing page cross-check on the test date. Free-tier sizes, character quotas, and trial credit amounts were confirmed against each vendor's public pricing page on 2026-09-16. Where a vendor's page is ambiguous, we say so explicitly instead of inventing a number.
- Exclusion evidence. Vendors on the exclusion list were checked for a free API tier on the same date. ElevenLabs' API pricing (from five dollars per month), OpenAI TTS's pay-per-character model (no free tier), and MiniMax Audio's sales-contact requirement are all current as of the test date.
- No projected or rumored quotas. If a number could not be verified by a live response or an official page, it does not appear in this article. This is why some sections say "not publicly documented" instead of quoting a figure.
This protocol is repeated on every monthly refresh, so the ranking's evidence base ages gracefully rather than rotting silently.
Practical Integration Details Worth Knowing
Beyond choosing an option, these implementation details save hours in real projects.
Audio format selection. MP3 at 24 kHz mono is the sweet spot for voice content: roughly 28 kilobytes per minute of speech with no audible quality loss for spoken word. Use WAV only when downstream processing (pitch shifting, diarization) demands lossless input; the file size penalty is about ten times. OGG/Opus is worth considering for bandwidth-constrained mobile delivery.
Caching identical text. Voice prompts, error messages, and UI confirmations are synthesized from the same strings thousands of times. Hash the text plus voice plus speed parameters, cache the resulting audio, and you will typically eliminate sixty to eighty percent of TTS calls — which directly stretches any character-based free quota by the same factor.
Chunking long text. Most free tiers cap request size (commonly a few thousand characters). Split long documents at sentence boundaries rather than hard character limits, keep chunks under the cap with margin, and concatenate the audio outputs. Splitting mid-word produces audible clicks that no amount of post-processing fully removes.
Handling 429 gracefully. Implement exponential backoff with jitter, and read the reset header when present. A fixed sleep interval either wastes quota (too long) or hammers the service into longer bans (too short). For anonymous tiers like TTS.ai's, the practical answer is simply to register — the quota triples and the 429 wall moves far away.
Voice consistency across a product. Lock the voice name, speaking rate, and pitch parameters in a single configuration file. Different voices across onboarding, notifications, and error messages make a product feel assembled from spare parts, and migrating voices later means re-synthesizing every cached clip.
For a unified way to manage these parameters across multiple free providers, the Apishare gateway exposes aggregated models behind one OpenAI-compatible key, so your caching, chunking, and retry logic is written once and works everywhere.