Orchard vs AssemblyAI.

AssemblyAI's Universal model sits at $0.0020 per minute for async transcription — already one of the more aggressive per-minute rates among full-feature STT vendors. Orchard ships at $0.00042 per minute, 5× cheaper than AssemblyAI Universal and (as of 2026) the lowest per-minute STT rate of any production-grade API publishing public pricing.

Where the gap widens further: audio intelligence add-ons. AssemblyAI charges $0.06/hour for auto chapters, $0.06/hour for sentiment analysis, $0.06/hour for entity detection. Stack three of those onto a 30k-min/month workload and you're at $150/month total AssemblyAI bill vs $25 flat on Orchard Pro — and the Pro plan also covers your TTS + voice cloning workload.

AssemblyAI's strongest product differentiation is LeMUR — their LLM layer that runs on top of the transcript for summaries, Q&A, entity extraction, action items. It's well-engineered and deeply integrated. Orchard doesn't ship a direct equivalent; we keep the surface minimal (transcript in, transcript out) and expect callers to pipe the result into their own LLM (Claude, GPT-4, Llama, whatever) for downstream intelligence.

The tradeoff:

• LeMUR convenience: one API call, structured output, no separate LLM key to manage. Costs per-token on top of the transcription bill.
• DIY approach: you control the LLM, prompt, output schema, fallback strategy. Cheaper at scale (most teams already pay an OpenAI/Anthropic bill); locks in nothing. Adds ~10 lines of code.

Audio intelligence add-ons (auto chapters, sentiment, entity detection, PII redaction) are on our roadmap; today we don't ship them as packaged features. If you need them as a one-API experience right now, AssemblyAI is the right pick.

AssemblyAI Universal is a proprietary model trained from scratch; it benchmarks well on English (typically top-3 on public WER leaderboards). Orchard runs a tuned Whisper-large derivative (whisper.cpp + Core ML) — same model family as OpenAI's original Whisper API, with our optimizations for throughput and Spanish.

• English (clean): Universal wins by ~1-2% absolute WER on US-accent corpora.
• English (accented): Roughly parity. Whisper's broader pre-training corpus helps with accents.
• Spanish (neutral): Parity within ±1.5% WER.
• Spanish (rioplatense): Orchard wins by ~4% absolute — we ship a fine-tune for porteño speech.
• Multispeaker diar: Both ship native diarization. Quality is comparable on standard podcasts.

Both vendors are async-first with optional streaming. AssemblyAI ships a real-time WebSocket today; Orchard's equivalent is still on the roadmap, with the sync HTTP endpoint covering most voice-agent use cases at ~150 ms p50.

• Real-time partial transcripts: AssemblyAI wins via WebSocket streaming.
• Sync HTTP (short utterance): Roughly parity at ~150 ms.
• Batch (long-form): Orchard shards across the cluster — 60 min audio in ~90 s wall (40× real-time). AssemblyAI batch is comparable.
• Webhooks on completion: both support callback URLs.

AssemblyAI is STT-first and doesn't ship a TTS or voice cloning product. Orchard ships three on the same balance:

AssemblyAI's SDK is well-documented but proprietary. Orchard mirrors the OpenAI Whisper API shape so if you wrap STT behind your own service layer, migration is an env var change. Side by side:

// AssemblyAI
const client = new AssemblyAI({ apiKey: process.env.AAI_KEY });
const transcript = await client.transcripts.transcribe({
  audio: "https://example.com/podcast.mp3",
  speaker_labels: true,
});

// Orchard (OpenAI SDK, drop-in)
const client = new OpenAI({
  baseURL: "https://api.orchardrun.com/v1",
  apiKey:  process.env.ORCHARD_API_KEY,
});
const transcription = await client.audio.transcriptions.create({
  file:  fs.createReadStream("podcast.mp3"),
  model: "whisper-1",
});

// For diar + async: POST /v1/transcriptions/upload
//   form fields: file, diarize=true, webhook_url=https://...

The async + diar + webhook combo is documented at /docs#async and /docs#diarization.

Three scenarios where AssemblyAI is the better pick today:

• LeMUR is core to your product. If you ship a feature that depends on the LLM-on-transcript workflow as a single API call (meeting summarizer, voice assistant memory, call analytics with action items), the cognitive cost of swapping to "transcript + your own LLM call" can outweigh the price delta.
• You need audio intelligence add-ons as packaged features. Auto chapters, sentiment analysis, entity detection, PII redaction — these are first-class on AssemblyAI today. They're on our roadmap but not shipped.
• Real-time streaming voice agents. AssemblyAI's WebSocket streaming handles partial transcripts at sub-300 ms latency. Until our streaming endpoint ships, that's a real gap.

For everything else — high-volume batch transcription, multilingual workloads, anything Spanish-heavy, anything that benefits from TTS or voice cloning on the same bill, anyone minimizing infrastructure cost per minute — Orchard is the economically obvious choice at $0.00042/min, the cheapest production-grade STT rate on the market.