Speech Technology News

Speech Technology News: What Changed in Voice AI This Autumn

Speech technology news moved in three directions over the past seven weeks. Voice agents began listening while they talk, speech models split into specialized and on-device versions, and courts started to say what a voice is worth legally. Below is what shipped, what the companies claim, and what remains unproven. It is a snapshot of recent releases, not a full market survey. Related coverage lives in our technology section.

Voice agents that listen while they speak

On September 10, OpenAI released GPT-Live-1 in its API. OpenAI describes it as bringing ChatGPT’s full-duplex voice conversations to developers, with custom voices and telephony support for phone-call agents. Full duplex matters because the agent keeps listening while it speaks, so users can talk without waiting for a pause, unlike turn-based systems where only one side transmits at a time.

OpenAI says GPT-Live-1 improves on its earlier real-time model by 30 percentage points on Full Duplex Bench. That is the vendor’s own evaluation, so independent testing matters before anyone builds on it. Its announcement has the details.

Text-to-speech competes on emotion and speed.

On September 28, ElevenLabs launched Eleven v4, which it calls its most emotive text-to-speech model, plus a low-latency Turbo variant for voice agents. The company reports a median inference latency of around 100 milliseconds for Turbo, excluding network and application delays. It also says Eleven v4 ranked first on a September 2026 Artificial Analysis leaderboard.

Two practical notes. Voice clones created before the launch need retraining to work well with v4. And a leaderboard position is a snapshot. Rankings in this category shift quickly, so treat any “best model” claim as dated.

Speech recognition gets specialized and private

Several recognition releases landed on October 14, and a regional one followed today:

DateReleaseWhat the company claims
Oct 14Speechmatics Oak91.1% medical-term recall across 11 languages, versus an average of 74.3% for eight other providers.
Oct 14Tether QVAC TranscribePsyA 254-million-parameter multilingual model that runs entirely on-device, so audio never goes to cloud servers.
Oct 28Gnani AI Prisma v3.0Nine Indian languages plus Indian English, trained on Indian telephone audio, with four transcription modes.

All three comparisons come from the companies’ own announcements. Speechmatics’ figure is a vendor benchmark that competitors haven’t independently reproduced, as far as I could find. The pattern is still useful. The market is shifting from one general model toward models tuned for medicine, local languages, or privacy. If you use dictation or transcription for work, test any tool on your own audio. Accents, background noise, and jargon change results a lot. Our apps and software coverage is the place to compare consumer options.

Courts begin to protect the human voice.

The most consequential non-product story came from Japan. On September 30, the Tokyo District Court held that a voice, like a portrait, can be protected under publicity rights, the first such finding by a Japanese court. The case involved anime voice actor Kenjiro Tsuda and TikTok videos he said used an AI clone of his voice.

The details limit how far the ruling reaches. The court found that publicity rights cover a voice used purely for its commercial appeal, and Tsuda won in part. It dismissed his request to delete the videos because they had already been removed, and it did not rule on whether the AI voice actually copied his. TikTok said it would study the judgment before deciding its next steps. The ruling confirms voices can be protected. It doesn’t say when an imitation crosses the line, so it’s guidance more than a bright-line rule.

It lands as cloning gets easier. ElevenLabs says Eleven v4 can clone a voice from as little as 10 seconds of audio. Anyone producing synthetic voices commercially should get written consent from the speaker. The same applies to tools that imitate a famous voice, such as a Stephen Hawking voice generator.

What to take from this

  • Building voice agents: try full-duplex models against your own call scenarios, and check latency claims under real network conditions.
  • Using transcription: run a short test on your own recordings, and ask whether audio leaves your device if privacy matters.
  • Working with synthetic voices: get documented permission for any real person’s voice, whatever the law currently says in your country.

The next signals to watch are independent benchmarks of these new models and any appeal in the Tokyo case.

Comments

No comments yet. Why don’t you start the discussion?

    Leave a Reply

    Your email address will not be published. Required fields are marked *