Google AI Edge Eloquent: Free Offline AI Dictation (2026)

On June 12, 2026, Google released AI Edge Eloquent as a free offline dictation component inside the Google AI Edge SDK version 1.0.0. The model ran entirely on device. It required no internet connection, no Gemini API key, and no Google account after the initial 210 MB download. Developers could embed the dictation engine in Android, Linux, and Windows apps under the Apache 2.0 license. Google published benchmarks on the same day. The model supported 27 languages, handled punctuation and speaker labels, and posted a 4.3 percent word error rate on clean speech. On a Pixel 9, latency stayed under 120 milliseconds for short utterances. The announcement came from the Google AI team via the official developer blog and a GitHub release notes page.

The release hit two groups immediately. Free-tier developers who had been rationing Gemini API speech credits gained a local option with no metered cost. Privacy-sensitive apps in healthcare, legal, and field services gained a way to keep audio on device. Google had tightened its Gemini API free tier earlier in 2026. Some developers saw their speech-to-text access moved behind paid compute quotas. AI Edge Eloquent bypassed that billing path. It was not a cloud replacement. It was an edge model built to run on mid-range phones and laptops. The company said the on-device model was free for production use. Optional cloud fallback would still use standard Gemini pricing. That split mattered for teams already burned by AI free tier limits.

Why it mattered became clear in the competitive context. Google had spent months cutting cloud model prices. The company had already reduced Gemini Pro costs and tightened access to free tier resources. OpenAI and Anthropic had responded with usage-based billing changes and credit pool overhauls. AI Edge Eloquent extended that pressure to on-device inference. It was a direct answer to OpenAI Whisper, which required self-hosting or API fees for most offline use. Apple Dictation worked offline only for a limited set of languages on recent devices. Google undercut both by shipping a free model with an open license. The move sat inside a broader Google AI price cut strategy that had already made OpenAI and Anthropic nervous. The Stanford HAI 2026 AI Index noted that on-device inference models reduced cloud dependency for privacy-focused teams. For developers, the signal was plain: cloud speech was no longer the default.

For end users, the practical effect was simple. Dictation kept working in airplane mode. It did not pause for a network round trip. It did not leak audio to a remote server. The model was not a toy. Google trained it on 1.2 million hours of multilingual speech. It used a 90 million parameter transformer encoder that ran inside a 210 MB quantized package. The company said quantization reduced memory use by 38 percent compared to the full model with less than a 0.4 point accuracy drop. That mattered for phones with 4 GB of RAM. Teams could ship free dictation without adding a cloud bill. The release also landed inside a broader push for free AI models with no API costs.

How Do the Top Options Compare?

Benchmarks are vendor-reported except where noted. Offline language support varies by device and model version. Google AI Edge Eloquent accuracy was measured on LibriSpeech clean and noisy sets.

1. Google AI Edge Eloquent , Free offline dictation in production apps

Google AI Edge Eloquent was the headline release of June 12, 2026. It shipped in Google AI Edge SDK 1.0.0 under the Apache 2.0 license. The model ran on-device and needed no API key after initial download. Developers accessed it through a C++ and Kotlin API. The package was 210 MB in its default quantized form. It covered 27 languages at launch, including English, Spanish, Mandarin, Hindi, Arabic, and Portuguese. Google said the model handled punctuation, capitalization, and speaker diarization. On clean speech, measured word error rate was 4.3 percent. Under noisy conditions at 10 dB SNR, the rate rose to 7.1 percent. That was still better than Google’s previous on-device speech model by 2.8 points.

Why it mattered for pricing: the model was free for production use. No per-minute charge. No cloud round trip. No quota reset. Google had tightened its Gemini API free tier earlier in 2026. Teams that relied on cloud speech had watched free tier changes push some workloads behind paywalls. Eloquent gave those teams an escape hatch. It also fit the broader trend toward free AI models with no API costs or subscriptions.

For privacy, the difference was structural. Audio stayed on the device. No transcription logs reached Google servers unless a developer opted into cloud fallback. That mattered for HIPAA-adjacent workflows, legal dictation, and field inspections. Google said the model could run on mid-range Android phones with 4 GB of RAM. Peak memory use was 380 MB during inference. The company reported that quantization cut memory use by 38 percent from the full model. Accuracy loss was 0.4 points on the standard LibriSpeech test set.

Key strengths:

• ✅ Free for commercial and personal use under Apache 2.0
• ✅ Runs fully offline with no API key or cloud round trip
• ✅ Covers 27 languages with punctuation and speaker labels
• ✅ Sub-120 ms latency on Pixel 9 and 380 MB peak memory
• ✅ Open license allows local model modification
• ❌ Offline model accuracy trails cloud Gemini speech on noisy audio
• ❌ 210 MB download may be large for low-storage devices
• ❌ Cloud fallback costs standard Gemini API rates

Who it’s for: Developers who need private, zero-cost dictation in Android, Linux, or Windows apps.

2. Gemini API Speech-to-Text , High-accuracy cloud transcription at scale

Google’s cloud speech offering continued to serve large batch jobs and live captioning in 2026. It handled more languages and more complex audio than the on-device Eloquent model. Google had changed the Gemini API free tier earlier in the year. The free tier still included 60 minutes of speech per month, but overage required a paid plan. Standard pricing was $0.004 per 15 seconds. That worked out to $0.96 per hour of audio for standard models. Pro models cost more.

The cloud model posted a 3.1 percent word error rate on clean speech and 5.4 percent under noisy conditions. It supported 125 languages and dialects. It also handled medical and technical vocabulary better than the 210 MB edge model. But every call sent audio to Google servers. That made it less attractive for privacy-first apps.

Google’s own price cuts had made cloud speech cheaper than in 2025. The company reduced standard speech pricing by 22 percent in April 2026. Still, the edge release made the cloud option look like a paid fallback rather than a default. Developers could route short utterances to Eloquent and send long files to Gemini. That hybrid pattern fit the broader AI pricing changes in June 2026.

Key strengths:

• ✅ 125 plus language support with pro-grade medical and technical vocabulary
• ✅ Higher accuracy for noisy or accented audio
• ✅ Batch processing for long files
• ✅ Simple API with streaming and speaker diarization
• ❌ Audio leaves the device and incurs metered charges
• ❌ Free tier capped at 60 minutes per month
• ❌ Requires internet and API key

Who it’s for: Teams that need maximum accuracy or batch transcription and can accept cloud processing.

3. OpenAI Whisper , Open-source speech recognition research and self-hosting

OpenAI Whisper remained the most popular open-source speech model in 2026. The model family was available under the MIT license on Hugging Face and GitHub. Developers could run Whisper locally, but it was not optimized for mobile. The large-v3 model required 1.5 GB of VRAM or more for real-time use. Latency on CPU was often above 800 milliseconds for short clips. On a GPU, latency dropped to around 150 milliseconds, but that hardware was not in most phones.

Whisper supported 99 languages. Its word error rate on clean English was 4.0 percent. That was close to Google’s Eloquent. But Whisper’s larger memory footprint and slower CPU performance made on-device deployment difficult. Self-hosting was free, but the cost of engineering and hardware was real.

OpenAI also offered Whisper through its API. Cloud pricing was $0.006 per minute. That was $0.36 per hour, cheaper than Google’s standard speech rate in some cases. But the API did not offer a free tier for transcription. Developers paid from the first minute. That context made Google’s free offline release notable. For a full breakdown, see AI API free tiers and limits.

Key strengths:

• ✅ Open MIT license with many fine-tuned community variants
• ✅ 99 languages and strong multilingual performance
• ✅ Self-hosting avoids per-minute fees
• ✅ Good accuracy for English and clean audio
• ❌ CPU latency can exceed 800 ms, making real-time mobile use hard
• ❌ Model sizes from 39 MB to 1.5 GB plus for large variants
• ❌ API transcription has no free tier

Who it’s for: Researchers and self-hosters who can manage GPU workloads and model tuning.

4. Apple Dictation , Basic built-in dictation on iPhone, iPad, and Mac

Apple Dictation was already free on Apple devices in 2026. It worked offline for a limited set of languages on newer iPhones and Macs. The feature supported more than 30 languages online, but offline language support was smaller. Apple did not publish detailed word error rates or latency numbers. Users reported good performance for short messages but weaker handling of technical terms.

Apple tied dictation to its own operating systems. Developers could not embed Apple Dictation in Android or Windows apps. That limited its use for cross-platform products. Apple also kept the model closed. There was no standalone license for third-party apps.

Compared with Google AI Edge Eloquent, Apple Dictation required no download and no setup. That was its main advantage. But it did not offer an open SDK for external developers. The free offline dictation race in 2026 had clear platform boundaries. Google’s move put pressure on Apple to extend offline language support. For context on free AI pricing changes in June 2026, see the full report. Free AI News may earn a commission if you sign up for a paid plan through links on this site. This does not affect our reporting.

Key strengths:

• ✅ Built into iOS, iPadOS, and macOS with no download
• ✅ Free for end users
• ✅ Works with systemwide text fields and apps
• ✅ Offline support on recent Apple devices for limited languages
• ❌ No third-party SDK for Android or Windows developers
• ❌ Offline language support is smaller than Google’s 27-language launch
• ❌ Apple does not publish accuracy benchmarks

Who it’s for: Apple device users who want basic dictation without installing anything.

Frequently Asked Questions

What Should You Remember?

• Free offline dictation: Google released AI Edge Eloquent on June 12, 2026 with no API key or cloud fee.
• On-device privacy: Audio stays local unless developers opt into cloud fallback.
• Accuracy benchmark: 4.3 percent word error rate on clean speech and 7.1 percent at 10 dB SNR.
• Language support: 27 languages at launch, more packs promised later in 2026.
• Hardware limits: Runs on 4 GB RAM devices with 380 MB peak memory and a 210 MB download.
• Competitive pressure: Google undercut OpenAI Whisper and Apple Dictation on free offline pricing.
• Developer license: Apache 2.0 allows commercial use without paying Google.

Free AI News is an independent editorial publication. Information about AI pricing, free-tier limits, and features changes frequently and may become outdated. Always verify current details through the vendor’s official pages. Affiliate links may earn a commission at no cost to you, and never affect our reporting.