7 Best AI Audio Tools & APIs for Developers (2026)

The best AI audio tools for developers in 2026 are Deepgram for streaming speech-to-text, AssemblyAI for transcription plus audio intelligence, ElevenLabs for premium text-to-speech, the OpenAI Realtime API for all-in-one voice agents, Krisp for on-device noise cancellation and turn-taking, Dolby’s media APIs for post-production, and Stable Audio 3.0 or Suno for music. Picking between them comes down to two constraints most buyers check last: latency and licensing. Get those right and the rest of the shortlist falls out on its own. Get them wrong and you ship a voice agent that lags half a second behind every reply, or a music feature your lawyer kills the week before launch. We’ve wired all seven tools below into production products, so this is the list we actually reach for, with the 2026 prices, the SDK coverage, and the exact spot where each one breaks.

Updated October 2026: gpt-realtime-2.1 and 2.1-mini prices, Stable Audio 3.0 and Suno v6, Krisp VIVA 2.0, re-checked Deepgram, AssemblyAI and ElevenLabs rates, cost math re-run.

Why trust this shortlist

Fora Soft has built software since 2005: 250+ shipped projects, 50+ in-house specialists, with a long specialisation in real-time media: WebRTC, streaming, and the audio pipelines that sit underneath them. Audio is not a side quest for us. We run these vendors in live products right now: Deepgram and AssemblyAI for captioning, ElevenLabs and OpenAI for voice agents, Krisp on the mic for call-heavy apps, and Dolby’s media APIs for post-production.

The clearest proof is TransLinguist, an AI interpretation platform we built that transcribes and translates speech across 75+ languages using Deepgram, Google Cloud Speech, and Speechmatics behind one MediaSoup/WebRTC pipeline. It won the UK’s NHS national framework for language services and has been featured twice in Slator. So when we say a tool wins or breaks at a given slot, it’s from wiring it up, not from reading a vendor deck. If you want that read on your own stack, our AI integration team does exactly this.

How to evaluate an AI audio tool in 2026

Before you compare vendors, pin down six axes. They decide the shortlist. Skip this and you’ll benchmark a real-time engine against a batch one, or pay a premium for on-device inference you don’t need.

1. Latency class. Under 300 ms for conversation, under 2 s for live captioning, 5–60 s for async. Each class has a different winner.

2. Accuracy on your audio. Word error rate on clean English says nothing about a user on a subway platform. Test on your own corpus before you sign.

3. Licensing and output rights. For TTS and music generation: who owns the output, is commercial use allowed, is training-data indemnity included?

4. SDK coverage. Web, iOS, Android, server. Partial coverage means shipping two SDKs or writing your own bridge.

5. Compliance posture. HIPAA for healthcare, SOC 2 Type II for enterprise, a GDPR DPA for EU users, and clear retention terms for everything.

6. Unit economics at scale. The headline per-minute price matters less than concurrency tiers, volume discounts, and egress at your projected usage.

Figure 1. The audio stack in six layers. Each of the seven tools owns a slot; a real product usually wires together three or four of them.

AssemblyAI — transcription plus audio intelligence

What it is. A speech-to-text API with batch and streaming modes, speaker diarization, and a bundle of audio-intelligence features: summarization, sentiment, chapters, topic detection — you’d otherwise build with a second LLM call.

Why it matters in 2026. The Universal model line holds up well on noisy, accented English, and the extras save real engineering time on podcast, meeting, and legal/medical workflows. If your feature is “turn this recording into structured notes,” this is the shortest path there.

October 2026 pricing. Pre-recorded Universal-2 is $0.15/hour and Universal-3.5 Pro is $0.21/hour; streaming is $0.15/hour on Universal-Streaming (English or multilingual) and $0.45/hour on Universal-3.6 Pro Realtime. Billing is per second, and speech-understanding add-ons bill separately at default effort: sentiment +$0.02/hr, summarization +$0.02/hr, speaker identification +$0.02/hr, entity detection +$0.08/hr, topic detection +$0.15/hr. The free tier covers up to 185 hours of pre-recorded and 333 hours of streaming audio. Rates checked October 15, 2026 on the AssemblyAI pricing page.

Where it breaks. Turn on three or four add-ons and the “$0.15/hour” headline doubles. And for sub-300 ms turn-taking it isn’t the tightest option — that’s the next tool.

Deepgram — lowest-latency streaming speech-to-text

What it is. A streaming STT platform built for tight loops, with the Nova-3 model family, on-prem and VPC deployment, and in 2025 a dedicated voice-agent model, Flux, plus a Voice Agent API.

Why it matters in 2026. Voice agents and live interpreters live or die on turn-taking latency. Deepgram consistently delivers the tightest streaming loop, and Flux adds turn-awareness so your agent knows when the caller actually stopped talking. The on-prem option is the only viable path for some regulated buyers.

October 2026 pricing. Pay-as-you-go streaming lists Nova-3 Monolingual and Flux English at $0.0077/min regular (limited-time promo rates of $0.0048 and $0.0065 were live on October 15, 2026) and Nova-3 Multilingual at $0.0092/min. Pre-recorded Nova-3 is $0.0043/min. The bundled Voice Agent API starts at $0.075/min, and Aura-2 text-to-speech is $0.030 per 1,000 characters. New accounts get a $200 credit, and the Growth plan cuts every rate. Details on the Deepgram pricing page.

Where it breaks. The audio-intelligence layer is thinner than AssemblyAI’s, so if you want chapters and summaries out of the box you’ll bolt on your own LLM step. When no single engine covers your accuracy, diarization and on-prem needs at once, that routing layer is what our custom speech-to-text development team builds.

ElevenLabs — premium multilingual text-to-speech

What it is. A text-to-speech and voice-cloning platform with 70+ languages on Eleven v3, professional and instant clones, an emotion-control API, and a low-latency Flash model around 75 ms time-to-first-byte.

Why it matters in 2026. ElevenLabs voices are hard to distinguish from human audio in blind tests, and the multilingual v3 model closed the accent gap. For audiobooks, dubbing, and premium voice agents it’s the default. The consent-verification flow also speeds up enterprise procurement.

October 2026 pricing. The TTS API lists $0.08 per 1,000 characters on Multilingual v2 and v3 and $0.04 on Flash/Turbo. Plans run Starter $6/mo, Creator $22/mo, Pro $99/mo, Scale $299/mo and Business $990/mo. The same API now sells Scribe v2 speech-to-text at $0.22/hour and the Eleven Music API at $0.15 per generated minute. See the ElevenLabs API pricing page. For the wider TTS field, we compared six engines in our synthetic voice libraries guide.

Where it breaks. Cost scales with characters, so a chatty agent that reads long responses gets pricey fast. And voice IDs can be deprecated, which is a real operational risk if you hard-code one.

OpenAI Realtime API and Whisper — conversational agents

What it is. OpenAI’s Realtime API — the flagship is now gpt-realtime-2.1 (July 2026), an update of GPT-Realtime-2, which brought GPT-5-class reasoning to voice in May 2026 — bundles speech-in, reasoning, and speech-out into one WebRTC/WebSocket session. Whisper stays the open-weight workhorse for batch transcription, and GPT-Realtime-Whisper ($0.017/min) now streams captions as the speaker talks.

Why it matters in 2026. Realtime collapses the STT→LLM→TTS triple-hop into one call, so you stop synchronising three vendors, and OpenAI says the July 2026 update cut p95 latency across its Realtime voice models by at least 25%. For a greenfield voice agent it’s the fastest route to a working demo.

October 2026 pricing. gpt-realtime-2.1 audio is billed at $32 per 1M input tokens ($0.40 cached) and $64 per 1M output tokens; gpt-realtime-2.1-mini is $10 and $20. Audio runs about 600 tokens per minute of user speech and 1,200 per minute of model speech, so a mixed minute of talk starts near $0.04 before the conversation history is re-billed on every turn. With that replay and prompt caching working, plan on $0.06–0.11 per minute on the flagship and $0.02–0.05 on mini. Our OpenAI Realtime API pricing breakdown has the per-turn math. Whisper’s open weights are free to self-host, and whisper.cpp runs on-device.

Where it breaks. Convenience has a price. At scale, one bundled vendor costs multiples of a specialist stack (we do the math below), and you trade away fine-grained control of each stage.

Krisp — on-device noise cancellation and turn-taking

What it is. A noise-cancellation, voice-isolation, and accent-conversion SDK that runs entirely on-device, now with VIVA turn-taking models that predict when a speaker is about to finish and whether a mid-response noise is a real interruption.

Why it matters in 2026. The quality of every downstream step (STT, agents, recordings) is capped by how clean the mic signal is. Krisp cleans it before the network hop, which cuts both bandwidth and downstream API cost. On the turn-taking side, VIVA 2.0 (May 2026) added Turn Prediction v3, an audio-only Interrupt Prediction model and Voice Isolation v3. Krisp says VIVA now carries more than 12 billion minutes of voice-agent traffic a year across 130+ voice AI products, and that platforms using it report 3.5× better turn-taking accuracy. That matters because clumsy turn detection is what makes voice agents feel robotic.

2026 pricing. The developer Voice AI SDK is quote-based, typically per active user or per concurrent seat. Krisp ships the VIVA 2.0 models at no extra charge on existing VIVA pricing.

Where it breaks. It’s a pre-processor, not a transcriber or a voice — it makes the other tools better, it doesn’t replace them. And SDK pricing is opaque until you talk to sales. For the accuracy math behind denoising, see our guide to speech recognition in noisy environments.

Dolby.io Media APIs — mastering, enhance, diagnose

What it is. A set of REST APIs for broadcast-grade post-processing: Enhance (denoise plus EQ), Music Mastering (loudness normalization), Diagnose, and Analyze. Since the 2025 rebrand to Dolby OptiView, the new developer docs cover Player, Live, Real-time Streaming and Ads; the media-processing APIs sit in Dolby’s legacy documentation.

Why it matters in 2026. For podcast platforms, UGC apps, and creator tools, one Enhance pass on an amateur recording lifts perceived quality a full tier — the difference between “phone call” and “studio.” It does in one API call what used to need an audio engineer.

2026 pricing. As of October 2026 the OptiView site lists no public per-minute rate for the media APIs, so get a written quote before you build on them. If you want a priced, self-serve alternative, Auphonic’s API runs the same leveling, denoise and loudness pass, with API access on every plan, including a free tier of 2 hours of audio a month.

Where it breaks. It’s async, file-based post-processing, not a real-time filter — don’t reach for it inside a live call. And the brand churn means older tutorials point at dead URLs, so budget a little integration archaeology and review the dependency every year.

Suno and Stable Audio 3.0 — music and sound generation

What they are. Two generative-music options. Suno produces full songs with vocals and leads on quality; its v6 model shipped on September 9, 2026, built with Warner Music Group, BMG and Believe after Warner settled its lawsuit against Suno in November 2025. Stable Audio 3.0 from Stability AI (May 2026) generates instrumental tracks up to 6 minutes 20 seconds on its Medium and Large models, and Stability says every 3.0 model is trained on fully licensed data.

Why they matter in 2026. Music is the newest slot in the audio stack and the riskiest on rights. Suno sounds best but still has no official public API — access is the web app or third-party resellers, and Suno has published nothing on API access for v6. Stable Audio 3.0 Large ships through the Stability AI API and as a self-hosted enterprise deployment, and ElevenLabs’ Eleven Music API ($0.15 per generated minute) is a third API option. Licensed training data is far easier to defend to a commercial buyer.

October 2026 pricing. Suno Pro is $10/mo (2,500 credits, about 500 songs) and Premier $30/mo (10,000 credits); only paid plans carry commercial rights. Stability has not published a per-request price for Stable Audio 3.0 that we could verify; Stable Audio 2.5 still runs about $0.20 per output file through API resellers. Under Stability’s Community License you own your outputs; organisations above $1M in annual revenue need an enterprise license, which adds legal indemnification.

Where it breaks. Copyright. In Part 2 of its AI report (January 2025), the US Copyright Office said work generated entirely by AI from a prompt generally isn’t protectable — writing the prompt doesn’t make you the author. Read the Copyright Office AI guidance before you build a business on generated tracks. If a song idea is coming from our music-app work, we route product teams to Stable Audio for anything shipping commercially.

The seven AI audio tools compared

One view of the whole shortlist. Read the “where it breaks” column as hard as the price column — that’s the one procurement forgets.

Figure 2. The same seven tools sorted by latency class. Choose the band your workload lives in first; the vendor list narrows itself.

Mini-case: building TransLinguist’s real-time translation stack

The problem. TransLinguist needed live, speech-to-speech interpretation inside video calls — a speaker in one language, listeners hearing another within a couple of seconds, across 75+ languages, reliable enough to win a national healthcare contract. No single vendor covered every language pair at the accuracy bar.

The stack we built. We put the STT layer behind an abstraction and routed each language to its best engine: Deepgram, Google Cloud Speech, or Speechmatics — then fed a translation and text-to-speech step, all riding one MediaSoup/WebRTC pipeline with FFmpeg for recording. The router meant a weak language pair on one engine could fall back to another without touching the app.

The outcome. The platform now offers AI speech-to-speech translation in 16+ languages and captioning in 22, serves a marketplace of 30,000+ interpreters, won the UK NHS national framework for language services, and has been featured twice in Slator. The multi-engine routing is the part we’d reuse on any serious audio product — and it’s the same instinct behind this whole guide: no single tool wins every slot. Want a similar assessment of your stack?

Build vs buy: the MAU threshold

The honest rule we give clients: integrate until roughly 5,000 monthly active users, then reconsider. Below that, hosted vendors are cheaper and safer than anything in-house. Above it, fine-tuned and partly self-hosted pipelines start to pay off — especially if your audio is distinctive: medical jargon, a specific accent, a domain vocabulary the big models fumble. That’s usually where our voice assistant development team comes in: we put Deepgram, ElevenLabs, OpenAI Realtime or Cartesia behind one abstraction, so you can swap engines as prices move.

Four buyer profiles and what we actually recommend:

Pre-seed / MVP. OpenAI Realtime for a voice agent, AssemblyAI for podcast features. One-vendor simplicity beats a cost optimisation you don’t have time for yet.

Seed to Series A. Split the stack (Deepgram plus ElevenLabs plus Krisp) and lock volume discounts before traffic arrives.

Growth (10k+ MAU, noisy or accented audio). Fine-tune Whisper on your corpus, keep Deepgram as fallback, self-host denoising where the latency budget allows.

Enterprise / regulated. On-prem Deepgram or a custom Whisper deployment, ElevenLabs via private endpoint, and a signed DPA with every vendor.

Cost math: a 10,000-MAU voice app

Assume 10,000 MAU, an average of 6 minutes of conversation per user per month, and 60% of caller-side audio needing denoise. Here’s the arithmetic, at 2026 list prices:

Talk volume: 10,000 × 6 min = 60,000 minutes = 1,000 hours/month.

Deepgram STT: 60,000 min × $0.0077/min (Flux or Nova-3 streaming at the regular pay-as-you-go rate) ≈ $462/mo.

ElevenLabs TTS: the agent speaks about half the talk time: roughly 15.6 million characters on the Flash model at $0.04 per 1,000 ≈ $624/mo.

Krisp denoise: $0.10 per active user (assumed SDK rate) × 6,000 noisy users = $600/mo.

Specialist total ≈ $1,686/mo for the voice layers, with Krisp at an assumed rate and before the text-LLM bill your agent pays separately. The same talk volume on OpenAI gpt-realtime-2.1, reasoning included, lands near $3,600/mo at the low end of its $0.06–0.11/min planning range and up to $6,600 at the high end — about 2.1× the voice-layer bill. The gap, minus your LLM spend, is the price of one-vendor simplicity. gpt-realtime-2.1-mini changes the picture: at $0.02–0.05/min it lands at $1,200–3,000/mo, in the same range as the specialist stack plus an LLM, if its reasoning is good enough for your agent. At 10k MAU the flagship can be worth paying for; by 100k it isn’t.

Figure 3. The specialist stack versus one-vendor gpt-realtime-2.1 at 10,000 MAU (October 2026 list prices). Convenience roughly doubles the voice-layer bill.

Licensing and rights: the part procurement rejects

Accuracy gets the demo approved. Licensing gets the launch approved. Three clauses trip up more audio products than any benchmark, so check them before you fall for a voice.

Voice-clone consent. If you clone a real person’s voice, you need documented consent from that person. ElevenLabs enforces a verification flow for exactly this reason. Skipping it is a lawsuit, not a shortcut.

Music output rights and copyright. Suno grants commercial rights on paid tiers, but the US Copyright Office position is that fully AI-generated tracks may not be protectable at all — you can use them, but you may not own them. Stable Audio’s licensed-data training is the more defensible base for anything you ship commercially.

Speech-data retention. For healthcare and finance, where the recording goes and how long it lives is the whole conversation. AssemblyAI, Deepgram, and OpenAI offer BAAs on the right tiers; Krisp runs client-side, which moves the question back to your own app. Get every term in writing.

5 integration pitfalls we’ve fixed

1. Denoising after STT instead of before. Cleaning audio after transcription removes audible hiss but can’t recover the accuracy you already lost. Put Krisp, or any denoiser, at the microphone boundary.

2. Billing per attempt, not per result. Most STT vendors bill for the open stream, not the words returned. A user who abandons mid-sentence still costs you. Add client-side voice-activity detection to close idle streams.

3. Hard-coding a single voice ID. Voices get deprecated with short notice. Put a “character → voice-id” map behind the call and keep a fallback so a deprecation is a config change, not an outage.

4. Ignoring codec negotiation. Opus at 48 kHz beats G.711 at 8 kHz by a wide margin on word error rate. If your WebRTC signalling quietly falls back to narrowband, your accuracy drops and no model can fix it.

5. Treating music generation as a feature, not a legal question. Teams ship generated tracks, then discover at the licensing review that they can’t own or clear them. Decide the rights posture before you wire the API.

KPIs: what to actually measure

Quality KPIs. Word error rate on your own audio (not the vendor’s demo set), mean opinion score for TTS naturalness, and signal-to-noise gain from the denoiser in decibels. Track them on real user recordings, sampled weekly.

Business KPIs. Cost per active minute, blended vendor spend per MAU, and the share of sessions that hit a paid feature. If cost-per-minute climbs faster than revenue-per-minute, the stack needs a rebalance.

Reliability KPIs. End-to-end turn latency at the 95th percentile, stream error rate, and failover success rate when a primary engine times out. A voice agent that’s fast on average but janky at p95 still feels broken.

When NOT to add AI audio tools

Honesty sells better than a feature list, so here’s the counter-case. If your audio is a one-off (a handful of files a week), a hosted consumer app beats any API integration on cost and effort. If your users are all on studio hardware in quiet rooms, a denoiser adds latency and spend for no measurable gain. And if you’re pre-product-market-fit, a bundled model like OpenAI Realtime gets you to a testable demo faster than a three-vendor stack you’ll rip out anyway once you learn what users need.

Add the specialist stack when the audio feature is core to the product and the volume justifies the integration — not before. Over-engineering the audio layer of an app nobody uses yet is the most common way we see teams burn a quarter.

A decision framework in five questions

Answer these in order and the shortlist writes itself.

Q1. What’s the latency budget? A conversational loop under half a second means Deepgram streaming STT plus a fast TTS, or OpenAI Realtime. Seconds are fine? AssemblyAI or Whisper open the field.

Q2. Is voice output part of the product? If yes, ElevenLabs for quality, or OpenAI Realtime if it’s bundled into the agent.

Q3. Are your users in noise? If yes, Krisp at the mic before anything else touches the signal.

Q4. Do you generate or just process audio? Generating music or SFX pulls in Stable Audio or Suno — and a licensing review.

Q5. What’s your MAU and regulatory surface? Under 5k and unregulated, integrate. Above that or regulated, plan for fine-tuning, on-prem, and DPAs. Still unsure after five questions? That’s the conversation we have on a scoping call.

Figure 4. What to integrate first, by product type. Start at the box that matches your product and add layers from there.

Frequently asked questions

What are the best AI audio tools for app development in 2026?

For most product teams the working set is Deepgram or AssemblyAI for speech-to-text, ElevenLabs for text-to-speech, Krisp for on-device denoise and turn-taking, Dolby’s media APIs for post-production, and Suno or Stable Audio 3.0 for music. OpenAI Realtime (gpt-realtime-2.1, or 2.1-mini on a budget) is the fastest way to a voice-agent MVP but costs more at scale.

Do I really need a denoiser if my STT is good?

Usually yes. Modern STT models handle noise only to a point, but every decibel of signal-to-noise improvement upstream lowers word error rate downstream — and cuts LLM cost if you chain STT into an agent. Put the denoiser before the STT engine, at the microphone, not after.

Can I run AI audio tools on-device in 2026?

STT and denoise, yes: whisper.cpp, Moonshine, and the Krisp SDK run on modern phones and laptops. Premium TTS at ElevenLabs quality is still cloud-only, though smaller on-device voices work for non-premium cases. Music generation is cloud-only for now.

Who owns music generated by Suno or Stable Audio?

It depends on the plan and the platform. Suno grants commercial usage rights on Pro and Premier; the free tier is personal-only. But the US Copyright Office holds that work generated entirely by AI generally isn’t copyrightable, so you may be able to use a track without being able to own it. Stable Audio 3.0’s licensed-data training is the safer base for commercial products. Read both terms in full and keep a lawyer in the loop.

What’s a realistic latency for a voice agent?

End-to-end — user stops speaking to agent starts speaking — of 250–500 ms is achievable with Deepgram plus an LLM plus ElevenLabs and solid WebRTC plumbing. OpenAI Realtime alone lands in the 200–400 ms band. Above one second, users start talking over the agent.

Is HIPAA a blocker for AI audio tools?

Not for this shortlist. AssemblyAI, Deepgram, and OpenAI offer BAAs on appropriate tiers, and ElevenLabs offers one on enterprise plans. Krisp runs client-side, so the HIPAA question shifts to your own app. Always confirm BAA scope in writing before you ship.

What about AWS Transcribe, Google, and Azure Speech?

They’re reasonable defaults if you’re already deep in one cloud and accept slightly lower accuracy and latency. For a dedicated audio product, the specialists win the benchmarks and the developer ergonomics. We use AWS Transcribe on projects like VocalViews where the app already lives on AWS.

How long does a real integration take?

A two-vendor voice agent (STT plus TTS) is 2–4 weeks to a working demo and 8–12 weeks to production with denoise, observability, and fallback paths. Music generation is a fast REST call, but the licensing and moderation review adds weeks. Podcast enhancement via Dolby’s API is the quickest, often under a week to a shipped feature.

What is the best AI model for audio processing?

There isn’t one model for every job. For speech-to-text, Deepgram Nova-3 or Flux wins on streaming latency and AssemblyAI Universal-3.5 Pro on transcription plus summaries; for speech output, ElevenLabs v3; for one model that hears, reasons and speaks, OpenAI gpt-realtime-2.1. Pick the latency class first, then test on your own recordings.

Which speech AI tools add speaker identification without building a model?

AssemblyAI sells speaker identification as an add-on at $0.02/hour on top of transcription (October 2026), next to diarization, sentiment and summaries. Deepgram returns diarized transcripts too. Both are an API parameter, so you don’t train or host anything.

What is the best speech enhancement SDK for developers?

For live audio, Krisp’s on-device SDK, which since VIVA 2.0 (May 2026) also covers turn prediction and voice isolation for voice agents. For uploaded files, an async API such as Auphonic or Dolby’s media APIs. Put the real-time denoiser at the microphone, before speech-to-text.

Which APIs generate audio in real time for web apps?

For speech, ElevenLabs Flash (about 75 ms to first byte by ElevenLabs’ figure), Deepgram Aura-2 ($0.030 per 1,000 characters) and OpenAI gpt-realtime-2.1 all stream audio back over WebSocket or WebRTC. Music is still request-and-wait: Stable Audio 3.0 and the Eleven Music API return a finished file, not a live stream.

What is the best SDK for building a live audio room app?

The room itself is a WebRTC problem: LiveKit, Agora or a mediasoup build carries the voices, and the AI tools above plug into that stream for captions, moderation or an agent. TransLinguist runs exactly this way, with AI interpretation on top of a MediaSoup pipeline.

Read next

Which AI audio tools should you actually ship?

There’s no single best tool — there’s a stack. For most teams shipping this year the path is Deepgram or AssemblyAI for speech-to-text, ElevenLabs for text-to-speech, Krisp for denoise and turn-taking, Dolby’s media APIs for post, and Stable Audio 3.0 or Suno if music is part of the product. OpenAI Realtime is the fastest route to a voice-agent MVP, at roughly 2.1× the per-minute cost of the specialist stack once you scale.

The order of decisions never changes: latency class first, licensing posture second, SDK coverage third, unit economics last. Run a two-week spike on your own audio before you sign anything annual — and if you want a second pair of eyes that has shipped this before, that’s us. For the deeper audio engineering behind these choices, our Audio for Video knowledge base goes layer by layer.

Sources & references: AssemblyAI, Deepgram and ElevenLabs pricing pages (checked October 15, 2026); OpenAI GPT-Realtime-2 (May 2026) and gpt-realtime-2.1 (July 2026) announcements; Krisp VIVA 2.0 (May 2026); Stability AI Stable Audio 3.0 (May 2026); Suno v6 (September 2026); Dolby OptiView docs; Auphonic pricing; US Copyright Office AI report, Part 2 (January 2025); IETF RFC 6716 (Opus); Fora Soft TransLinguist and VocalViews client projects.