Voice is now a production interface: what to use and what to watch in 2026

Speech synthesis (TTS): commercial and open-weight options

What is verified as of September 2026, and what you have to test yourself in Spanish and French.

What to test. Our rule: write 50 hard sentences per language (numbers, dates and amounts with local separators, tax IDs, proper names, acronyms, English terms), generate them with two or three vendors and run a blind listening test with five people from the target audience. A voice with a Spain accent can sound foreign in Bogotá, and Quebec French is not Paris French.

Transcription (STT): accuracy, diarization and migrations

OpenAI is renewing its line-up: GPT Transcribe (files) and GPT Live Transcribe (low latency) cost $0.017 per minute, the same as GPT-Realtime-Whisper, its streaming transcription model from May. whisper-1, gpt-4o-transcribe, gpt-4o-mini-transcribe and gpt-4o-transcribe-diarize leave the API on February 26, 2027. If you pay $0.006 per minute for legacy Whisper today, your cost almost triples when you migrate.

Google lists Gemini 3.5 Transcribe at about $0.005 per minute. ElevenLabs offers scribe_v2 (90+ languages), scribe_v2_realtime (about 150 ms, per the company) and scribe_v2_medical for clinical audio. Among open weights with clear licences: Whisper large-v3-turbo (about 809M parameters, MIT), NVIDIA Parakeet TDT 0.6B v3 (CC-BY-4.0, 25 European languages) and Mistral Voxtral Mini 3B (Apache 2.0, Spanish and French included).

Accuracy: measure on your own audio. A public leaderboard's WER does not tell you how your contact centre will fare. Test 8 kHz phone audio with real noise, regional accents and dictated numbers (ID numbers, tax IDs, licence plates): OpenAI specifically highlighted better alphanumeric recognition in GPT-Realtime-2.1 in July 2026. If you rely on diarization, confirm that the replacement for gpt-4o-transcribe-diarize covers your case before 2027.

Voice agents: speech-to-speech or cascade

Speech-to-speech. A single model takes audio in and answers with audio. GPT-Live-1 is full-duplex (it listens while speaking) and hands reasoning and tool use to a backend agent. It only runs on OpenAI's Live endpoint, and its rate limits are counted in concurrent sessions, from 25 at Tier 1 to 500 at Tier 5, with no free tier. Google positions Gemini 3.8 Live as the default Live API model for low-latency voice agents.

Cascade. Streaming STT, an LLM with tools and streaming TTS, each piece swappable. You gain control: you see the text of every turn, apply rules before anything is spoken and pick the best Spanish voice. You pay in latency and turn-taking engineering.

Our decision rule: if the conversation is open-ended and naturalness matters more than control, start with speech-to-speech. If there are transactions with personal data, business rules or audit needs, start with a cascade, or with a speech-to-speech model that delegates to a backend where you control tools and rules, which is what GPT-Live-1 proposes. More in AI agents in production.

Latency, turn-taking and interruptions

A budget per stage and five rules so the conversation doesn't sound robotic.

Barge-in and turn-taking

1. Keep listening while the agent speaks, with echo cancellation.
2. When the user interrupts, stop the audio and drop the unspoken part from the history: the model must know what the person heard.
3. Don't count an 'uh-huh' or a cough as an interruption; tune the threshold on real recordings.
4. Detect end of turn by meaning, not just silence: someone dictating a number pauses.
5. Read critical data back (ID numbers, amounts, dates) and confirm before acting.

AI music: models, lawsuits and licences

The market is moving from unlicensed training to licensed models, and courts are setting the pace.

Suno sums up the moment. Its v6 was built with Warner, BMG and Believe, and the company is retiring all earlier models. Sony and Universal, which did not sign, claim v6 "launders" past infringement because it was trained partly on earlier models' outputs.

In Europe, the Munich Regional Court ruled on July 31, 2026 that Suno infringed by training on GEMA works without licences and by reproducing works that match famous songs in melody, harmony and rhythm.

If your brand uses generated music

• Use plans with explicit commercial rights: on Suno, Pro and Premier downloads carry them, while free downloads are for personal use only.
• Keep evidence for every campaign asset: tool, plan, date, prompt and file.
• Don't ask to imitate a specific artist or a recognizable voice.
• Expect distribution friction: Deezer says more than 50% of new uploads in June 2026 were fully AI-generated and removes them from recommendations.

Voice cloning: consent, watermarks and fraud

Voice cloning is now a product feature: since v5.5 (March 2026), Suno lets you record your own voice and reuse it in songs. The rule is explicit consent: a contract stating which voice, for which uses, in which languages, for how long and how it is revoked.

Since August 2, 2026, AI Act Article 50 requires synthetic audio to be marked in a machine-detectable way and audio deepfakes to be disclosed; the Commission's code of practice asks for signed metadata plus a watermark. NIST concluded that no single provenance or detection technique is enough on its own.

Reference architecture and compliance: a Spanish-language voice agent

1. Entry: SIP or WebRTC telephony with an opening notice: you are talking to an AI, the call is recorded, why, and how to reach a human.
2. Engine: speech-to-speech or cascade, behind your own interface so you can switch vendors.
3. Voice: STT and TTS chosen with your 50-sentence test in Colombian Spanish, and a brand voice under contract.
4. Backend: an LLM with strict tools (CRM, orders, tickets), least privilege, spoken confirmation before any change, and defenses against prompt injection.
5. Handoff: to a human, with a summary, when the user asks or after two failed attempts; identity is verified with a factor other than voice.
6. Operations: p50 and p95 latency, a reviewed sample of transcripts, resolution rate without a human, and automatic deletion of recordings when the retention period ends.

Compliance

European Union. Since August 2, 2026, AI Act Article 50 requires telling people clearly, at the latest at the first interaction, that they are talking to an AI; fines reach €15 million or 3% of worldwide turnover, whichever is higher. A recording is personal data under the GDPR, and a voiceprint used to identify someone is special-category biometric data (Article 9), which in practice usually requires explicit consent. Using AI to infer your own employees' emotions has been prohibited since February 2025, except for medical or safety reasons.

Colombia. Law 1581 of 2012 requires prior, informed authorization that you can prove later (Article 9): store the notice and the acceptance with each call. Biometric data is sensitive (Articles 5 and 6) and needs explicit authorization, and sending audio to a processor abroad requires a data transmission contract (Decree 1074 of 2015) or, for a transfer, compliance with Article 26. There is no AI law in force yet, but Law 2502 of 2025 already makes AI impersonation an aggravating factor. Details in our 2026 regulatory map.