Introduction
Voice agents live or die on the telephone, yet most text-to-speech comparisons are recorded in quiet rooms with studio headphones. That matters because a caller on a mobile network hears heavily compressed 8 kHz audio, not full-bandwidth studio output. With OpenAI retiring its legacy speech models in January 2027, teams building phone agents face a forced migration whether they planned for one or not. This guide evaluates eight TTS providers specifically for real-world telephony, looking at native G.711 support, concurrency limits, interruption handling, and what actually survives the trip over a carrier line.
What Happened
On October 1, 2026, OpenAI announced that tts-1, tts-1-hd, and gpt-4o-mini-tts will be removed from its API on January 6, 2027. The suggested successor, gpt-realtime-2.1-mini, does not work with the /v1/audio/speech endpoint, so anyone relying on that path must switch providers or architectures this quarter. The broader market kept shipping: Cartesia released Sonic 3.6, Deepgram launched Flux TTS on a new /v2/speak socket, ElevenLabs introduced Eleven v4 and v4 Turbo, and AssemblyAI debuted AssemblyAI-TTS-1 on October 27, 2026.
Telephony audio is unforgiving. Carriers in North America and Japan typically use G.711 mu-law at 8 kHz, while much of the rest of the world uses A-law. Anything above 4 kHz is stripped away, and each sample is squeezed into 8 bits. That erases much of the detail that makes premium voices sound premium. Sibilants and fricatives, sounds like S and F, live mostly above 4 kHz, so spelling out email addresses or confirmation codes can become unintelligible. Time to first audio is also critical; silence on a call invites callers to talk over the agent.
Why This Matters
Seven of the eight APIs reviewed return G.711 natively, so basic telephony compatibility is no longer a meaningful differentiator. The real dividing line is concurrency: how many simultaneous calls an API can handle before queuing or rejecting requests. Some services offer unlimited sessions with only a per-minute session cap, while self-serve plans elsewhere top out at single digits. Barge-in behavior also varies widely. When a caller interrupts, the system must stop synthesis, flush the carrier buffer, and know exactly how much of the sentence the person heard. Only a few vendors expose that information explicitly.
OpenAI's shutdown adds urgency. Its speech endpoint returns MP3, Opus, AAC, FLAC, WAV, or 24 kHz PCM, never G.711, so developers must transcode audio themselves. After January 6, 2027, that endpoint disappears entirely, replaced by a different real-time architecture rather than a simple text-to-speech call.
Key Takeaways
- AssemblyAI-TTS-1: Built for high call volume with unlimited concurrency and native mu-law/A-law output at any supported rate, including 8 kHz. It offers per-word timings, a cancel function for barge-in, and instant voice cloning. The trade-off is a short six-language list and no SSML or custom lexicon.
- ElevenLabs: Strong voice range and cloning across more than 90 languages with the v4 family. Concurrency is plan-capped, and requests beyond the limit queue. Its v4 Turbo model runs on a separate Text to Dialogue socket that reserves one session per open connection.
- Cartesia Sonic 3.6: Designed for low-latency streaming with documented telephony output formats. Self-serve concurrency is limited to between 2 and 15 requests depending on the plan, making it better suited for pilots than large contact centers without an Enterprise agreement.
- Deepgram Flux TTS: Offers the most explicit barge-in contract, reporting exactly what text was spoken and what remains when interrupted. Regional streaming caps differ sharply, with far lower limits in the EU, Australia, and India than in North America.
- Rime: Useful for entity-heavy calls thanks to custom X-SAMPA lexicons and a spell function for codes and acronyms. Its Coda socket supports mu-law but not A-law, so transcoding is required outside mu-law regions.
- Azure AI Speech: Broad language coverage and enterprise-grade quotas, though limits are counted in transactions rather than calls, which can deplete quickly for chatty agents.
- Google Cloud TTS: Strong locale coverage with native G.711, but streaming requires Chirp 3: HD over gRPC, meaning teams must build a bridge from their carrier's WebSocket.
- OpenAI: Being phased out. Existing implementations must migrate before January 6, 2027, and handle mu-law transcoding on their own in the meantime.
Conclusion
Leaderboards that play full-bandwidth clips on headphones cannot predict how a voice agent will sound over a phone earpiece in a moving car. Benchmark sentences also rarely contain the confirmation codes, email addresses, and order numbers that dominate real agent traffic. The best way to choose is to synthesize your own most difficult prompts, push them through an 8 kHz mu-law codec, and measure what a listener can still recover. Concurrency caps, interruption handling, and model availability through 2027 will matter more than any preference score. Pick two vendors, test them against your actual call scripts, and trust the phone line over any leaderboard.
Disclosure: This article is sponsored content. HackerNoon's editorial team reviewed it for clarity and quality, but the views, claims, and comparisons are those of the sponsor, AssemblyAI.

Discussion
Join the conversation
Thoughtful reactions, questions, and follow-up ideas help shape the next story.