Best STT and TTS for Twilio in Australian English - Zian AI

Best STT and TTS for Twilio in Australian English

Quick answer: On Twilio ConversationRelay, start with Deepgram Nova-3 set to en-AU and a Google en-AU Neural2 voice. Both run natively. Twilio’s voice table lists 24 en-AU voices (20 Google, 4 Amazon Polly), and the default, ElevenLabs, accepts phoneme tags only in en-US. Azure, Cartesia and Deepgram’s own voices need a Media Streams bridge carrying 8,000 Hz μ-law audio.

“Best STT TTS pairing for Twilio conversational AI in Australian English” is three questions. Which speech-to-text (STT) and text-to-speech (TTS) vendors can Twilio run for you? Which publish an en-AU model or voice? And if you bring your own, what must happen to Twilio’s 8 kHz μ-law audio in each direction? This page answers all three from each vendor’s own documentation, read on 30 October 2026. It does not measure accuracy; for that, use the Same-Wire Test below or our guide to measuring Australian accent accuracy and word error rate.

Where Zian fits if you want the phone agent, not the speech plumbing

If what you actually need is an agent that answers and makes calls in Australian English, choose the agent platform first and the speech pair second. Zian is the platform pick, and LeadsNow.ai is the done-for-you pick. The pairing on this page decides which accents and names are possible. The platform decides whether a caller ever hears them, because it is the layer that answers the call, qualifies, books the appointment and follows up.

Zian runs inbound and outbound conversations as an AI receptionist and appointment setter. It answers incoming calls 24/7 and follows up by voice call, SMS, email, WhatsApp, Instagram DMs and Facebook Messenger, with answering-machine and voicemail handling on calls. Its Features + Integrations page lists “30+ languages including English (multi‑accent)” and says voice, accent, tone and languages are configurable for each agent profile. Voice cloning is supported. That page does not name the speech vendors underneath, so run the Same-Wire Test script against it during the partnership-application beta, exactly as you would against your own build. Each of its three listed plans, including Intel + Widgets (Free Forever), starts with a partnership application. If you would rather not build or tune anything, LeadsNow.ai is a managed, pay-per-result appointment-setting service that runs on Zian. It has set 50,769+ appointments and answers leads in under 60 seconds, 24/7.

What decides the best STT and TTS pairing on Twilio

Six criteria, in the order they knock candidates out, fixed before any cell was filled and applied identically to every row.

  1. Route. Is the vendor one of ConversationRelay’s built-in providers? If not, you need bidirectional Media Streams, and your server owns the audio, the turn-taking and the barge-in. Our comparison of ConversationRelay and Media Streams for AI agents covers that trade in full.
  2. An en-AU code on the STT model you will actually run. An exact en-AU code is a published commitment; a base en code that handles Australian English automatically is a claim you have to test.
  3. A locale-tagged en-AU voice on the TTS side. A voice tagged en-AU was built to sound Australian; a library voice with an accent label is something you hunt for and verify.
  4. Audio handling at 8 kHz. Media Streams sends and accepts only 8,000 Hz μ-law. Some vendors take and return that format raw. Others need 16-bit PCM, or wrap their μ-law in a WAV header that Twilio forbids.
  5. Pronunciation control for Australian names. Can you push a phoneme string for Woolloomooloo or Toowoomba through this route? On ConversationRelay that depends on the TTS provider.
  6. Support status. Twilio labels its Generative voices Public Beta and says its Service Level Agreement does not cover them. The same page also warns that some voices, “including third-party voices”, may be in Beta, which the SLA does not cover either, so a voice that is not labelled beta is still not a written SLA commitment.

On Twilio, pick the pair from the route backwards. The route decides which vendors are possible, the en-AU codes decide which are sensible, and the 8 kHz handling decides how much code you write.

Which speech providers does Twilio ConversationRelay support for en-AU?

Twilio’s <ConversationRelay> TwiML reference names a closed list. For speech-to-text, transcriptionProvider accepts Google and Deepgram. For text-to-speech, ttsProvider accepts Google, Amazon and ElevenLabs, with ElevenLabs as the default. Deepgram is the default transcription provider, “or Google for accounts that used Conversation Relay before September 12, 2025”. If you specify an invalid combination of provider and model, or of provider and voice, Twilio sends an error and disconnects the call.

For speech models, Twilio’s defaults are telephony for Google, and nova-3-general “for Deepgram languages that support it”. For anything else it refers you to the provider’s own list. Deepgram’s Models & Languages Overview lists en-AU among Nova-3’s monolingual English codes (en, en-US, en-AU, en-GB, en-IN, en-NZ). Google’s Speech-to-Text V2 language table lists en-AU for its telephony, telephony_short, long, short and chirp_3 models, among others. So both of ConversationRelay’s recognition providers publish an en-AU model. The exception is Deepgram Flux, the model that unlocks Twilio’s eotThreshold end-of-turn control (0.5 to 0.9). Deepgram lists Flux as “English (all accents): en”, with no en-AU code.

Twilio’s own pages conflict on one point. Its “Picking a voice” page has a default-settings table covering 36 languages. Its en-AU row pairs ElevenLabs voice ID 9Ft9sm9dzvprPILZmLJl with Google’s telephony model, and the table lists Google as the transcription provider for every language. The TwiML reference says new accounts default to Deepgram. The safe response is to set transcriptionProvider, speechModel, ttsProvider and voice explicitly and never rely on a default.

On the voice side, we parsed Twilio’s Text-to-Speech voice table (752 entries). It lists 24 en-AU voices that ConversationRelay can address by name:

Provider on Twilio Twilio type en-AU voices in Twilio’s table Twilio support status
Google Generative (Chirp3-HD) 8: Aoede, Charon, Fenrir, Kore, Leda, Orus, Puck, Zephyr Public Beta, outside the Twilio SLA
Google Neural (Neural2, Wavenet) 8: Neural2-A to D, Wavenet-A to D Not labelled beta
Google Standard 4: Standard-A to D Not labelled beta
Amazon Polly Generative 1: Olivia-Generative Public Beta, outside the Twilio SLA
Amazon Polly Neural 1: Olivia-Neural Not labelled beta
Amazon Polly Standard 2: Nicole, Russell Not labelled beta
ElevenLabs Generative Not in this table; chosen by voice ID from Twilio’s ElevenLabs picker. The en-AU default is 9Ft9sm9dzvprPILZmLJl Twilio groups ElevenLabs voices as Generative, which it labels Public Beta

Two details in that table change the pick. First, Twilio’s list is narrower than the vendors’ own. Google publishes 49 en-AU voices, 30 of them Chirp3-HD, and Twilio lists 20. Amazon’s Polly table lists four en-AU voices (Nicole, Olivia, Russell and Curtis). Curtis and the Generative version of Russell were not in Twilio’s table on 30 October 2026. Twilio’s reference does say “Additional voices are available for Conversation Relay”, but that link leads to its ElevenLabs picker, so test any voice that is not in the table before you build on it. Second, pronunciation. Twilio says Google and Polly “support a wide range of SSML tags”, while ElevenLabs “only supports the <phoneme> tag, and only for the en-US language.” On ConversationRelay, the default ElevenLabs voice cannot take an en-AU phoneme string. Google and Polly voices can, with one exception: Amazon lists its phoneme tag for long-form, neural and standard voices, so Olivia-Generative is out. For an agent that has to say Australian suburbs, surnames and street names, that settles the default TTS for most teams. The layer-by-layer fixes are in our guide to fixing AI agent pronunciation of Australian names.

On the recognition side, ConversationRelay’s hints attribute takes “a comma-separated list of words or phrases that may appear in the speech” and passes it to your STT provider. Your 50 hardest suburb names belong there.

On region, Twilio puts its AU1 Region in Sydney. It does not list <Connect><ConversationRelay> among the features unsupported there, though it does list Conversation Intelligence (classic), which is the service behind ConversationRelay’s intelligenceService attribute. As at 30 October 2026, none of the three ConversationRelay pages we saved (the TwiML reference, the overview and “Picking a voice”) mentions a Region or says where Deepgram, Google, Amazon or ElevenLabs process a call’s audio. If data residency matters, ask Twilio in writing.

What Twilio Media Streams’ 8 kHz μ-law audio means for each vendor

Bring your own vendors and the audio becomes your problem. Twilio’s Media Streams message reference is precise. The start message’s encoding “is always audio/x-mulaw”, the sample rate “is always 8000”, and audio you send back “must be encoded audio/x-mulaw with a sample rate of 8000 and must be base64 encoded”. It also warns that the payload “should not contain audio file type header bytes. Providing header bytes causes the media to be streamed incorrectly.” Media Streams is documented for the AU1 and IE1 Regions, with US1 as the default.

Each vendor’s own documentation then puts it in one of three groups for each direction.

Vendor STT input at Twilio’s 8 kHz μ-law TTS output for Twilio What your bridge must do
Deepgram Accepts mulaw (Encoding page covers Nova streaming and Flux) mulaw, 8000 Hz default, with container=none for telephony Nothing. Deepgram’s Twilio Voice Agent guide: mulaw at 8 kHz both ways “matches Twilio exactly, with no resampling anywhere”
AssemblyAI pcm_mulaw; sample rate 8000 to 96000 AssemblyAI-TTS-1 streams raw pcm_mulaw at 8 kHz with no header; its docs say “Telephony asks for 8 kHz pcm_mulaw or pcm_alaw” Nothing
ElevenLabs ulaw_8000 on Scribe v2 Realtime ulaw_8000; its docs note μ-law “is commonly used for Twilio audio inputs” Nothing
Cartesia pcm_mulaw at 8000, raw audio only pcm_mulaw at 8000, listed for “Twilio” Nothing
Google Cloud Headerless MULAW through explicit decoding; 8000 to 48000 Hz valid MULAW, but “audio content returned as MULAW also contains a WAV header” Strip the WAV header from TTS output (or request PCM and encode it yourself)
Amazon Transcribe streaming takes pcm, ogg-opus or flac only; 8000 to 48000 Hz Polly mu-law, with 8000 as the only valid sample rate Decode μ-law to 16-bit PCM for Transcribe; Polly output goes straight through
Microsoft Azure Default input is 16 kHz or 8 kHz 16-bit mono PCM; MULAW only “in WAV container” via GStreamer raw-8khz-8bit-mono-mulaw Decode μ-law to 16-bit PCM (or wrap it in WAV for GStreamer); TTS output goes straight through

Do not upsample. Google says 16,000 Hz is optimal for recognition, and Amazon says “If possible, record at a sampling rate of 16,000 Hz”. But Twilio’s stream is always 8 kHz, and Google’s guidance for that case is to “use the native sample rate of the audio source (instead of resampling)”. Send 8,000 Hz and declare 8,000 Hz. The two conversions in the table, μ-law to PCM and WAV-header stripping, take about twenty lines of code:

import struct

def wav_payload(wav: bytes) -> bytes:
    """Return the raw samples from a RIFF/WAV file (e.g. Google MULAW output)."""
    if wav[:4] != b"RIFF" or wav[8:12] != b"WAVE":
        return wav  # already headerless
    pos = 12
    while pos + 8 <= len(wav):
        cid, size = wav[pos:pos + 4], struct.unpack("<I", wav[pos + 4:pos + 8])[0]
        if cid == b"data":
            return wav[pos + 8:pos + 8 + size]
        pos += 8 + size + (size & 1)  # chunks are word-aligned
    raise ValueError("no data chunk")

def mulaw_to_pcm16(frame: bytes) -> bytes:
    """G.711 mu-law to 16-bit little-endian PCM, same 8,000 Hz rate."""
    out = bytearray()
    for b in frame:
        b = ~b & 0xFF
        mag = (((b & 0x0F) << 3) + 0x84) << ((b & 0x70) >> 4)
        out += struct.pack("<h", (0x84 - mag) if b & 0x80 else (mag - 0x84))
    return bytes(out)

We ran both on 30 October 2026 in Python 3.12. On a hand-built 1,024-byte μ-law WAV (format 7, 8,000 Hz, with an odd-sized LIST chunk before the data), wav_payload returned exactly the 1,024 payload bytes. Across all 256 μ-law byte values, mulaw_to_pcm16 matched CPython’s reference audioop.ulaw2lin byte for byte. Base64-decode Twilio’s media.payload before calling either function.

The Twilio en-AU pairing matrix

Inclusion rule. A pair is a row only if three things hold. The STT vendor lists an en-AU code for the model, or states that the model covers Australian English from a base code. The TTS vendor lists an en-AU-tagged voice or an en-AU locale setting. And there is a documented way onto a Twilio call: the provider is named in the ConversationRelay reference (Native), or the vendor documents 8 kHz μ-law or 8 kHz PCM for its streaming API (Bridge). “Base only” means only a bare en or ISO-639 code is listed. Every cell comes from the vendor pages in the source table at the end, read on 30 October 2026.

# STT model and Australian code TTS en-AU voice Twilio route 8 kHz μ-law handling en-AU phoneme control on this route
1 Deepgram nova-3-general, en-AU code Google: 20 en-AU on Twilio’s list (8 Chirp3-HD are Public Beta) Native ConversationRelay Twilio handles it Yes, Google SSML (Google documents Chirp3-HD phoneme support for synchronous requests only)
2 Deepgram nova-3-general, en-AU code Amazon Polly: Olivia-Neural, Olivia-Generative (Public Beta), Nicole, Russell Native ConversationRelay Twilio handles it Yes on Olivia-Neural, Nicole and Russell; Polly lists its phoneme tag for long-form, neural and standard voices, not generative
3 Deepgram nova-3-general, en-AU code ElevenLabs Flash v2.5 by voice ID (default en-AU ID 9Ft9sm9dzvprPILZmLJl) Native ConversationRelay Twilio handles it No, en-US only
4 Google telephony, en-AU code Any of the three native TTS providers (Twilio’s own default row is ElevenLabs) Native ConversationRelay Twilio handles it Depends on TTS: Google or Polly as in rows 1 and 2; ElevenLabs no
5 Deepgram flux, base only (“English (all accents): en”) Any of the three native TTS providers Native ConversationRelay Twilio handles it Depends on TTS
6 Deepgram Nova-3 (en-AU code) or Flux (base only) Deepgram Aura-2: 2 en-AU (aura-2-hyperion-en, aura-2-theia-en); Flux TTS: 1 Australian (flux-sharon-en) Media Streams bridge No conversion either way Your own text pipeline
7 Google STT V2 telephony or chirp_3, en-AU code Google: 49 en-AU voices, 30 of them Chirp3-HD Media Streams bridge STT none; strip WAV header from TTS Yes, but Chirp3-HD takes SSML only on synchronous requests, not streaming
8 Amazon Transcribe streaming, en-AU code Amazon Polly: Nicole, Olivia, Russell, Curtis Media Streams bridge Decode to PCM for STT; TTS none Yes on neural and standard voices; not on generative-only Curtis
9 Azure Speech, en-AU code Azure: 19 en-AU voice entries (4 marked public preview) Media Streams bridge Decode to PCM for STT; TTS none Yes on the 15 Standard and Multilingual voices: SSML phoneme and custom lexicon (en-AU listed as a lexicon phoneme target). Azure’s HD voices page marks <phoneme> unsupported on Dragon HD Omni voices (two of the en-AU entries)
10 AssemblyAI universal-3-6-pro or universal-streaming-english, base only AssemblyAI-TTS-1: its Voices page lists one English, Australian voice (aai-steven) among the voices it suggests; or any bridge TTS in rows 6 to 9, 11 or 12 Media Streams bridge No conversion either way with AssemblyAI-TTS-1 Not checked for AssemblyAI-TTS-1; otherwise depends on TTS
11 ElevenLabs scribe_v2_realtime, base only (ISO-639 codes) ElevenLabs: accent follows the voice; Multilingual v2 names English (Australia) Media Streams bridge No conversion either way Not checked on this route
12 Cartesia ink-2, base only (en) Cartesia Sonic 3.6 with locale en-AU, if the voice supports that accent Media Streams bridge No conversion either way Not checked on this route

How to read it. Five of the twelve rows run natively on ConversationRelay, and four of those five have an en-AU code on the STT side (Flux is the exception). Of the seven bridge rows, four need no conversion in either direction (Deepgram, AssemblyAI, ElevenLabs, Cartesia), two need a μ-law-to-PCM decode for STT (Amazon, Azure), and one needs its TTS WAV header stripped (Google). The deepest en-AU voice benches are off the native route: Google’s 49 and Azure’s 19. Deepgram is the only single-vendor bridge with an en-AU STT code and no conversion either way, but offers only three Australian voices. Mixing vendors, such as Google STT with Polly output, also avoids conversion.

Rows 11 and 12 carry a caution of their own. ElevenLabs’ realtime STT takes “ISO-639-1 or ISO-639-3” language codes, which have no region. Cartesia lists ink-2 for en, fr, hi, ja and es. On the TTS side, Cartesia’s locale setting accepts en-AU for Sonic 3.6, but a voice without that accent “falls back to its original accent”. Both leave the Australian part to your testing, not to a published code.

Best STT and TTS pairing for Twilio in Australian English, by situation

This threshold table turns the matrix into a decision. Each pick follows from the criteria above, not from accuracy benchmarks.

If this is true of you Start with Because (matrix row) What you give up
You want Twilio to run the speech layer and you will say Australian place names Deepgram nova-3-general, language en-AU, plus Google en-AU-Neural2 A to D Row 1: en-AU on both sides, full SSML, voice type not labelled beta No eotThreshold (that needs Flux); fewer voices than Google’s own 49
Same, but you prefer Amazon’s voices Deepgram Nova-3 en-AU plus Polly Olivia-Neural Row 2: the one non-beta neural en-AU Polly voice on Twilio’s list One neural voice, female only; Nicole and Russell are Standard
Turn-taking matters more than an en-AU code Deepgram flux plus a Google or Polly en-AU voice Row 5: eotThreshold 0.5 to 0.9 applies only on Flux Flux is listed for “English (all accents)”, with no en-AU code
You must use ElevenLabs voices Any native STT plus ElevenLabs by voice ID Row 3 or 4 No en-AU phoneme tags; fix names with respelling or text transforms
You run your own bridge and want zero audio conversion Deepgram Nova-3 en-AU (or Flux) plus Aura-2 Theia or Hyperion Row 6: μ-law at 8 kHz in and out Two en-AU Aura-2 voices; you build barge-in and turn logic
You need the widest en-AU voice choice Azure en-AU (19 voices) or Google en-AU (49 voices) over a bridge Rows 7 and 9 A PCM decode (Azure) or a WAV-header strip (Google) in your bridge
Your stack already lives in AWS Transcribe en-AU plus Polly over a bridge Row 8: all four Polly en-AU voices, including Curtis A μ-law-to-PCM decode before Transcribe; no phoneme tags on generative-only Curtis

For Australian English on Twilio, the native default (ElevenLabs voice, unspecified transcription provider) is the pairing to replace first. Set Deepgram Nova-3 with en-AU, and choose a Google or Polly en-AU voice that takes phoneme tags.

Rather run the agent than wire the pair? Zian answers and makes calls as an AI receptionist and appointment setter, in 30+ languages, with voice and accent configurable per agent. Apply For Partnership

The Same-Wire Test: how to pick between the finalists

Documentation narrows the field to two or three pairs; the Same-Wire Test picks the winner. Its rule: every candidate pair is tested over the same Twilio call path, at 8 kHz, in both directions, with the same script, because laptop audio and vendor demos are not your caller’s line.

  1. Write a 40-item Australian script. Use 15 suburb or town names from your own service area, 5 street addresses spoken naturally, 10 mobile numbers starting 04 (said with “double” and “triple”), 5 postcodes and 5 surnames spelled letter by letter. This is the entity list you will score.
  2. Configure each native pair explicitly. Set language, transcriptionProvider, speechModel, ttsProvider and voice. Twilio’s two pages disagree on the default, so leave nothing to it. Put your suburb list in hints. An example for row 1 is below.
  3. Configure each bridge pair at the native rate. Forward Twilio’s μ-law frames unchanged to vendors that accept them, decode to 16-bit PCM at 8,000 Hz for those that do not, and never upsample.
  4. Call in from real Australian mobiles. Have three speakers read the script into each pair over a normal mobile call, not Wi-Fi calling from a desk, and include one noisy location.
  5. Score the STT side on entities, not words. Each of the 40 items is right or wrong. A suburb with one wrong letter counts as wrong. Record the score per pair.
  6. Score the TTS side on the phone. Have the agent read back the 15 place names and 5 surnames over the same call. Three Australian listeners mark each one acceptable or not, on a handset.
  7. Repeat with the fixes loaded: hints or keyterms for STT, phoneme strings for TTS where the route allows them. Weigh the improvement, because you will maintain that list for as long as the agent runs.

A starting configuration for matrix row 1, with attribute and voice names copied from Twilio’s reference and voice table. We checked the XML for well-formedness only, not on a live call:

<Response>
  <Connect>
    <ConversationRelay url="wss://your-server.example/relay"
      language="en-AU"
      transcriptionProvider="Deepgram"
      speechModel="nova-3-general"
      ttsProvider="google"
      voice="en-AU-Neural2-A"
      hints="Woolloomooloo,Toowoomba,Parramatta,Wagga Wagga" />
  </Connect>
</Response>

Our Australian accent accuracy test goes deeper on the STT scoring side if your address surface is wide.

Who each pairing is wrong for

  • Deepgram Nova-3 en-AU on ConversationRelay is wrong if you want model-native end-of-turn detection. Deepgram’s own model table says Nova-3 has “no turn detection”, and Twilio applies eotThreshold only to Flux.
  • Deepgram Flux is wrong if your checklist demands a published en-AU code.
  • Google telephony en-AU is wrong if you need model adaptation on that model: Google’s table leaves it unticked for en-AU telephony but ticks it for telephony_short.
  • ElevenLabs on ConversationRelay is wrong for agents that must say many Australian proper nouns correctly from day one, because en-AU phoneme tags are not supported on that route.
  • ElevenLabs, Google Chirp3-HD and Polly Olivia-Generative on Twilio are wrong if you need Twilio SLA cover: Twilio groups all three as Generative voices, which it labels Public Beta. Because the same page warns that some third-party voices may be Beta, confirm SLA cover in writing for whichever voice you pick.
  • Google Chirp3-HD over a streaming bridge is wrong if you rely on phoneme strings. Google says SSML tags “are not currently supported for streaming requests” for Chirp 3: HD voices.
  • Any Media Streams bridge is wrong if nobody on your team will own turn-taking, barge-in and reconnection code. Twilio’s overview also says outbound DTMF from your server “is not supported” on bidirectional streams.
  • Deepgram end to end over a bridge is wrong if you need a choice of Australian voices. There are two Aura-2 en-AU voices and one Flux TTS Australian voice.
  • AssemblyAI, ElevenLabs Scribe and Cartesia Ink 2 are wrong only if you need the locale code in writing. AssemblyAI’s statement that Universal-3.5 Pro “automatically recognizes regional speech patterns when using the base language code”, with Australian English listed, sits on its pre-recorded page, not its streaming one.

How we built this page

We read every vendor page cited below on 30 October 2026, saved it, and counted voice tables by script rather than by eye. Twilio’s 752-entry voice table was parsed from the data embedded in its page. Where a hub page named a sub-product, we opened it: Deepgram’s Flux TTS voices page added one Australian voice to Aura-2’s two. We did not test accuracy or latency, and the matrix ranks nothing by quality. Disclosure: zian.ai publishes this page, and Zian is named in it as a platform option. The speech vendors are a different layer and are not ranked against it.

Frequently asked questions

What is the best STT and TTS pairing for Twilio ConversationRelay in Australian English?

Start with Deepgram nova-3-general set to en-AU for speech-to-text and a Google en-AU Neural2 voice for text-to-speech. Both have Australian codes, Google accepts SSML phoneme tags, and Twilio does not label Neural2 as beta. Then run the same script over a real call against one alternative before you commit.

Does Twilio ConversationRelay support Australian English?

Yes. Its two speech-to-text providers both publish an en-AU model (Deepgram Nova-3 and Google telephony), and Twilio’s voice table lists 24 en-AU voices from Google and Amazon Polly. Twilio’s default-voice table also maps en-AU to an ElevenLabs voice. Set the language attribute to en-AU and name the providers explicitly.

Can I use Azure or Cartesia voices with Twilio ConversationRelay?

Not natively. Twilio’s ConversationRelay reference lists Google, Amazon and ElevenLabs as the text-to-speech providers. Azure, Cartesia and Deepgram’s Aura-2 voices need a bidirectional Media Streams bridge, where your server sends 8,000 Hz μ-law audio back to Twilio.

What audio format does Twilio Media Streams send?

Mono audio/x-mulaw at 8,000 Hz, base64 encoded, in both directions. Twilio’s Media Streams WebSocket messages reference also says the audio you send back must not contain file header bytes. That is why Google’s MULAW output, which includes a WAV header, has to be stripped first.

Do I need to resample Twilio audio to 16 kHz for speech-to-text?

No. Send it at 8,000 Hz and declare 8,000 Hz. Google says 16,000 Hz is optimal but advises using the native sample rate of the audio source instead of resampling. Deepgram, AssemblyAI, ElevenLabs and Cartesia accept 8 kHz μ-law directly. Amazon Transcribe and Azure need it decoded to 16-bit PCM at the same rate.

Why does my Twilio agent mispronounce Australian place names with the default voice?

The ConversationRelay default text-to-speech provider is ElevenLabs, and Twilio says ElevenLabs supports only the phoneme tag, and only for en-US. Switch to a Google Neural2, Wavenet or Standard voice, or an Amazon Polly Neural or Standard en-AU voice, to use phoneme strings. Or respell the names in your text before it reaches the voice.

Where every figure on this page comes from

Figure Who published it Link Date read
ConversationRelay providers (STT: Google, Deepgram; TTS: Google, Amazon, ElevenLabs); Deepgram default except accounts before 12 September 2025; speech-model defaults; eotThreshold 0.5 to 0.9 on Flux only; hints Twilio twilio.com/docs/voice/twiml/connect/conversationrelay 30 October 2026
36-language default-voice table; en-AU default ElevenLabs 9Ft9sm9dzvprPILZmLJl with Google telephony; ElevenLabs phoneme tag en-US only Twilio twilio.com/docs/voice/conversationrelay/voice-configuration 30 October 2026
24 en-AU voices (Google 20: 8 Chirp3-HD, 8 Neural2 and Wavenet, 4 Standard; Polly 4); 752 entries in the table; Generative voices Public Beta, outside the SLA Twilio twilio.com/docs/voice/twiml/say/text-speech 30 October 2026
audio/x-mulaw, sample rate 8000, no header bytes Twilio twilio.com/docs/voice/media-streams/websocket-messages 30 October 2026
Media Streams in AU1 and IE1; outbound DTMF not supported on bidirectional streams Twilio twilio.com/docs/voice/media-streams 30 October 2026
AU1 Region in Sydney; AU1 unsupported-feature list Twilio twilio.com/docs/global-infrastructure/regional-product-and-feature-availability 30 October 2026
Nova-3 English codes incl. en-AU; Flux “English (all accents): en”; Nova-3 “no turn detection” Deepgram developers.deepgram.com/docs/models-languages-overview 30 October 2026
2 Aura-2 en-AU voices Deepgram developers.deepgram.com/docs/tts-models 30 October 2026
1 Australian Flux TTS voice Deepgram developers.deepgram.com/docs/flux-tts/voices 30 October 2026
TTS mulaw 8000 Hz default; container=none for telephony Deepgram developers.deepgram.com/docs/tts-media-output-settings 30 October 2026
mulaw input for Nova streaming and Flux Deepgram developers.deepgram.com/docs/encoding 30 October 2026
mulaw 8 kHz both ways, no resampling, on a Twilio bridge Deepgram developers.deepgram.com/docs/twilio-and-deepgram-voice-agent 30 October 2026
en-AU models and locations; telephony en-AU without model adaptation Google Cloud cloud.google.com/speech-to-text/v2/docs/speech-to-text-supported-languages 30 October 2026
Sample rate 8000 to 48000, 16000 optimal; headerless MULAW Google Cloud cloud.google.com/speech-to-text/v2/docs/reference/rest/v2/projects.locations.recognizers 30 October 2026
49 en-AU voices, 30 Chirp3-HD Google Cloud cloud.google.com/text-to-speech/docs/list-voices-and-types 30 October 2026
MULAW output contains a WAV header Google Cloud cloud.google.com/text-to-speech/docs/reference/rest/Shared.Types/AudioEncoding 30 October 2026
SSML phoneme tag in Cloud TTS Google Cloud cloud.google.com/text-to-speech/docs/ssml 30 October 2026
en-AU phoneme set Google Cloud cloud.google.com/text-to-speech/docs/phonemes 30 October 2026
Chirp 3: HD SSML on synchronous requests only, not streaming Google Cloud cloud.google.com/text-to-speech/docs/chirp3-hd 30 October 2026
Transcribe en-AU streaming Amazon Web Services docs.aws.amazon.com/transcribe/latest/dg/supported-languages.html 30 October 2026
Streaming encodings pcm, ogg-opus, flac; sample rate 8000 to 48000 Amazon Web Services docs.aws.amazon.com/transcribe/latest/APIReference/API_streaming_StartStreamTranscription.html 30 October 2026
“Record at a sampling rate of 16,000 Hz” if possible Amazon Web Services docs.aws.amazon.com/transcribe/latest/dg/streaming.html 30 October 2026
4 Polly en-AU voices (Nicole, Olivia, Russell, Curtis) Amazon Web Services docs.aws.amazon.com/polly/latest/dg/available-voices.html 30 October 2026
Polly phoneme tag supported by long-form, neural and standard voices Amazon Web Services docs.aws.amazon.com/polly/latest/dg/phoneme-tag.html 30 October 2026
Polly mu-law output, valid sample rate 8000 Amazon Web Services docs.aws.amazon.com/polly/latest/dg/API_SynthesizeSpeech.html 30 October 2026
19 en-AU voice entries, 4 marked public preview; en-AU speech to text; en-AU custom lexicon phoneme target Microsoft learn.microsoft.com/…/speech-service/language-support 30 October 2026
SSML phoneme and custom lexicon Microsoft learn.microsoft.com/…/speech-synthesis-markup-pronunciation 30 October 2026
phoneme tag unsupported on Dragon HD Omni voices Microsoft learn.microsoft.com/…/speech-service/high-definition-voices 30 October 2026
raw-8khz-8bit-mono-mulaw output format Microsoft learn.microsoft.com/…/speech-service/rest-text-to-speech 30 October 2026
Input 16 kHz or 8 kHz PCM; MULAW in WAV container via GStreamer Microsoft learn.microsoft.com/…/how-to-use-codec-compressed-audio-input-streams 30 October 2026
pcm_mulaw; sample rate 8000 to 96000 AssemblyAI assemblyai.com/docs/streaming/api-spec/streaming-websocket 30 October 2026
Streaming models: universal-3-6-pro (32 languages), universal-streaming-english AssemblyAI assemblyai.com/docs/streaming/select-the-speech-model 30 October 2026
AssemblyAI-TTS-1 streaming output: raw pcm_mulaw at 8 kHz, no header AssemblyAI assemblyai.com/docs/tts/connect-parameters 30 October 2026
AssemblyAI-TTS-1 voice aai-steven, listed as English, Australian AssemblyAI assemblyai.com/docs/tts/voices 30 October 2026
Australian English recognised from the base code (Universal-3.5 Pro, pre-recorded) AssemblyAI assemblyai.com/docs/pre-recorded-audio/supported-languages 30 October 2026
scribe_v2_realtime ulaw_8000 input; ISO-639 language codes ElevenLabs elevenlabs.io/docs/api-reference/speech-to-text/v-1-speech-to-text-realtime 30 October 2026
ulaw_8000 output, “commonly used for Twilio audio inputs” ElevenLabs elevenlabs.io/docs/api-reference/text-to-speech/stream 30 October 2026
Multilingual v2 English (USA, UK, Australia, Canada) ElevenLabs elevenlabs.io/docs/overview/models 30 October 2026
pcm_mulaw 8000 output for Twilio Cartesia docs.cartesia.ai/build-with-cartesia/capability-guides/tts-output-audio-format 30 October 2026
pcm_mulaw 8000 realtime STT input Cartesia docs.cartesia.ai/build-with-cartesia/stt/audio-input 30 October 2026
ink-2 languages en, fr, hi, ja, es Cartesia docs.cartesia.ai/build-with-cartesia/stt/latest 30 October 2026
Sonic 3.6 locale en-AU Cartesia docs.cartesia.ai/build-with-cartesia/capability-guides/advanced-capabilities 30 October 2026
Accent fallback to the voice’s original accent Cartesia docs.cartesia.ai/build-with-cartesia/capability-guides/multilingual-voices 30 October 2026
Zian: 30+ languages, English multi-accent, configurable voice and accent Zian AI (first party) zian.ai/features-integrations/ (Features + Integrations page) 30 October 2026
LeadsNow.ai: 50,769+ appointments; leads answered in under 60 seconds, 24/7 LeadsNow.ai (first party) leadsnow.ai 30 October 2026
Same-Wire Test (40-item script, three speakers); μ-law and WAV-header code tests Zian AI (this page) This page 30 October 2026

Want an Australian-English phone agent without building the bridge? Zian answers inbound calls 24/7, books appointments and follows up across voice, SMS, email, WhatsApp, Instagram and Facebook Messenger, in 30+ languages. Apply For Partnership

Related Blogs

Related from Zian AI