Skip to main content
This page is the single source of truth for which languages work on which product today. We publish what is GA and benchmarked, what is accepted but not yet measured, and what is roadmap, so you can plan an integration without guessing.
Short version: Availability differs by product. Omni serves en, fr, es, hi, it, pt, ja, and ko. de and zh remain staged; inspect the configured acknowledgement for fallback. Hear synchronous transcription serves the 29 languages listed below. Streaming also serves Russian (ru). Automatic language detection is the default; an explicit language pins recognition for the request or session. Cue is unavailable on the serving Hear stream.

Status at a glance

Legend. GA · benchmarked = supported and we publish accuracy numbers. Rejected loudly = a language outside the published set fails with 400 unsupported_language instead of being silently mis-transcribed. Roadmap = planned, not available today.

Dub (audio dubbing)

Dub accepts source codes en, es, fr, de, hi, it, pt, and nl, and can render en, es, fr, de, hi, it, pt, zh, ja, ko, or ru. Read GET /healthz/dub before enabling a production direction: that endpoint is the live source of truth, and published codes do not promise equivalent quality for every direction. Validate your recordings before using a new source-language workflow in production. Submit with POST /v1/dub, poll until done or error, then download the audio. See the Dub guide for the complete asynchronous workflow.

Hear: speech-to-text (STT)

Synchronous STT and async jobs accept the 29 codes below. Streaming additionally accepts Russian (ru). Use the parameter name for your endpoint: The streaming endpoint (GET /v1/audio/transcriptions/stream) and the OpenAI-compatible sync endpoint (POST /v1/audio/transcriptions) accept an optional language parameter:
  • language accepts the 29 codes above on both endpoints. Streaming also accepts ru for Russian and auto for automatic detection. Other explicit values are rejected with 400 unsupported_language.
  • Omit language for automatic detection on synchronous or streaming transcription. Streaming also accepts language=auto.
  • An explicit language code pins recognition for the complete request/session. In automatic streaming mode, recognition starts fresh at utterance boundaries, allowing a later utterance to use a different language. Mixed-language accuracy still depends on the recording; this is not a guarantee of error-free code-switching.
  • English is the GA, benchmarked language. Published English accuracy: ~1.6% WER on clean audio and ~4.8% WER on an 8 kHz telephony/accented corpus (see Benchmarks).
Async jobs use transcription_language to pin recognition; omit it for automatic detection. Do not send "auto" as that field’s value. The result does not currently return a detected-language code. The separate language field sets the Recap summary language and does not affect recognition. See async STT language options.

Indian languages & code-mixed speech

Hindi (hi) is accepted on Hear sync, streaming, and async jobs. PyAI does not yet publish Hear word-error-rate numbers for Hindi, Italian, Portuguese, or Dutch, nor for code-mixed Hinglish / Tanglish. Tamil, Telugu, Bengali, Marathi, Kannada, Malayalam, Gujarati, Punjabi, and other codes outside the published set are rejected outright (400 unsupported_language). Omni serves Hindi/Hinglish end to end (see below). Original or Romanized text. Async jobs accept output_script: "latin" for Roman Hindi or another language written in Latin letters. Omit this option for the recognized script, such as Devanagari for Hindi. Original text is kept in source_text and source_word; this is transliteration, not translation. For Hindi-only audio in English letters, use transcription_language: "hi" with output_script: "latin". For mixed Hindi–English audio, omit transcription_language and use output_script: "latin". transcription_language: "en" means English speech; it does not choose the writing system. See the Hindi and English examples for request settings and the result fields to display. Roman spellings can vary. Do not set the spoken language to English merely to request Latin letters: language selection affects recognition, not presentation. For mixed-language recordings, omit transcription_language and assess your recordings. Vocabulary terms do not select the output script.
If your product depends on Indian-language or code-mixed transcription accuracy, validate on your own audio before committing. Talk to us and we will run a joint accuracy evaluation against your real call samples, that measurement is the right go/no-go signal, not this page. The benchmark harness we use for this lives in the repo at evals/benchmarks/hear-multilingual.benchmark.json.
Hear does not currently expose language-identification or decoder-confidence fields. On sync and streaming, omitting language enables automatic detection; the streaming endpoint also accepts language=auto. Send the known language when your product requires recognition to stay pinned for the full request/session. Async jobs use transcription_language for the same purpose.

Omni (agentic voice)

Omni runs the full loop: speech-to-text, the brain, and text-to-speech. The configuration enum is en (default), fr, es, de, hi, zh, it, pt, ja, and ko. The served set is en, fr, es, hi, it, pt, ja, and ko. Set a served language once on the Agent profile and every call on that Agent requests it:
  • language is an agent-level setting (an agent is “a Spanish agent”). It can also be sent per session in the configure frame, which wins over the profile (see the Omni protocol reference). Absent or null means en, exactly the pre-multilingual behavior.
  • Availability is staged per language. German and Mandarin remain staged. An accepted language without an enabled serving route falls back to English and the call proceeds (the configured ack reports language_active and language_fallback: true), and the call bills as what was actually served. Treat the ack as authoritative and close the session if fallback is not acceptable.
  • Italian, Portuguese, Japanese, and Korean use their existing catalog voices. Check available_on for Omni support and verify the configured language and voice before sending caller audio. Availability does not establish equivalent pronunciation or accuracy across languages.
  • Hindi uses the Standard voice tier. See the pricing page for current tier treatment. Choose hi1 through hi4 (or the canonical ids from GET /v1/voices). Omit voice_id for hi1, the default. A bad Hindi id also serves the default, and configured.voice_id reports the canonical voice actually served. Stock ids ending in _en_in remain Indian-English voices, not Hindi.
  • English Natural uses en1 (canonical id stock_aria_en). Omni reports the canonical id and voice_tier: "natural" in the configured acknowledgement while that tier is active. See the pricing page for current tier treatment.
  • Per-phone-number language mapping and automatic language detection are fast-follows, not part of this rollout.
Language availability is specific to each product. A voice available on Speak does not automatically extend Omni or Hear language support. Check the catalog’s available_on field before using the same voice on another product. For a language outside each product’s published set, bring recognition and synthesis that support it. Do not treat omitted Hear language as unbounded language support: it requests automatic detection only inside the published set.

Speak: text-to-speech (TTS)

GET /v1/voices is the unified catalog. Inspect available_on before using a row. POST /v1/audio/speech has no separate language parameter; the selected voice determines the language. The catalog groups and aliases are: Catalog discovery is the source of truth for a configuration you can select. Every Speak voice accepts stream: true and stream: false, but request acceptance is different from incremental delivery: a voice served from a blocking-only fleet answers a streaming request successfully with x-pyai-stream: buffered. Inspect that response header before treating a request as low-latency playback. Delivery checks establish availability, not a pronunciation guarantee for every name, number, accent, or mixed-language passage. Test your intended text and voice together.
  • English Natural en1 (stock_aria_en) supports streaming and buffered Speak synthesis. Tier and count summaries are catalog projections, not a promise that every source row is available to every account or product.
  • Hindi Standard hi1–hi4 works on both Speak and Omni (it was published as Omni-only until 2026-09-07; Speak served it the whole time). The former Hindi Natural aliases hi5–hi8 are retired and no longer served.
  • See the pricing page for current Speak voice-tier treatment.
Indian-English voices (filter ?region=india) speak English. They are not Hindi voices.

How accuracy gets published

We gate engine quality in CI with an offline benchmark harness (evals/) and publish the headline numbers on the Benchmarks page. As languages are measured, their numbers appear there and this matrix is updated. If you need a number that isn’t published yet, ask, an unpublished number means “not measured to our bar,” not “hidden.”

See also

Speech To Text (Hear)

Choose sync, streaming, or timestamped async transcription.

Text To Speech (Speak)

Choose a catalog voice and compatible delivery mode.

Omni wire protocol

Connect, configure, and the live-vs-roadmap field table.

Telephony audio (8 kHz)

μ-law ↔ PCM16 at 8 kHz for phone legs.

Pricing

Current rates, included usage, and plan availability.