Speechify puts the phone requirement in one line: Twilio and LiveKit SIP expect 8 kHz mu-law or 16 kHz PCM. A text to speech API that returns neither leaves a developer to convert every clip before it reaches a caller.
Two terms come up in every output format list. A sample rate is how many measurements of the sound are stored each second, so 8 kHz means 8,000 a second. An encoding is how each measurement is written down, and mu-law and PCM are two encodings.
What are mu-law and PCM in text to speech?
PCM stores raw samples, while mu-law packs each sample into 8 bits using G.711, the telephony codec standard, and Google describes it as "8-bit samples that compand 14-bit audio samples".
OpenAI gives a concrete PCM definition: raw samples at 24 kHz, 16-bit signed, little-endian, without a file header. Bland describes the pcm output on its OpenAI-compatible endpoint the same way, as raw 24 kHz signed 16-bit little-endian mono audio, and names its native encoding pcm_s16le. Cartesia describes its mu-law option as "8-bit compressed" for "North American / Japanese telephony (G.711μ), Twilio". ElevenLabs says mu-law is "commonly used for Twilio audio inputs".
Arithmetic on those published formats shows the size gap. Mono mu-law at 8 kHz is 8,000 bytes a second, since each sample takes one byte. Mono 16-bit PCM at 8 kHz is 16,000 bytes a second, and the same PCM at 48 kHz is 96,000.
Which text to speech APIs return mu-law audio for phone calls?
Bland Speech, Amazon Polly, Azure, Google, ElevenLabs, Cartesia, Deepgram, and Inworld offer mu-law output, while OpenAI's speech endpoint returns mp3, opus, aac, flac, wav, and pcm, and Fish Audio returns mp3, wav, pcm, and opus, as of November 2026.
What these rows mean
- Encodings offered: the output formats each company names for its API. Google also offers MULAW and ALAW encodings.
- Sample rates offered: the rates each company names for those formats. Bland's mu-law is 8 kHz only, and Bland's PCM default is 48 kHz.
- OpenAI: the Realtime API accepts
audio/pcmuoutput, which is mu-law, while the/v1/audio/speechendpoint returns mp3, opus, aac, flac, wav, and pcm. OpenAI saystts-1,tts-1-hd, andgpt-4o-mini-ttsshut down on 6 January 2027.
Several companies say outright what the phone formats are for. Deepgram says "Requesting mulaw at 8 kHz matches Twilio's Media Streams format exactly", Rime says to set a sampling rate of 8000 "to match the telephony stream and shrink payloads", and Kalpa Labs names 8 kHz for telephony and 48 kHz for WebRTC.
What audio formats does Bland Speech return for a phone call?
Bland Speech returns either 16-bit PCM at 8000, 16000, 24000, 44100, or 48000 Hz, or 8-bit mu-law at 8000 Hz, and Bland describes the mu-law option as "For telephony".
A developer picks the format in the request. Bland's reference shows the audio block for a phone-ready call as:
"audio": { "encoding": "mulaw", "sample_rate": 8000 }
The default is different: output comes back at 48000 Hz, and Bland Speech v3 renders natively at 48 kHz, so Bland says requesting 48 kHz avoids a resample, which means an 8 kHz request involves one. Both formats stream over HTTP and a WebSocket, which Bland's text to speech API covers.
Pipecat maintains a Bland service and LiveKit offers a Python plugin for voice agents built on those frameworks.
Should a phone voice agent request mu-law or PCM?
A phone voice agent on a carrier or SIP leg that expects 8 kHz mu-law should request mu-law at 8 kHz, and an agent that talks to a browser should request PCM at a higher rate.
Mu-law at 8 kHz matches what Speechify, Deepgram, and Cartesia say Twilio and North American telephony take, and at 8,000 bytes a second it is smaller than any 16-bit PCM option above. PCM at 16 kHz is the other rate Speechify names for Twilio and LiveKit SIP. A higher rate such as 48 kHz fits where audio does not pass through a phone line, which is Kalpa Labs's WebRTC example. Bland's mulaw option is 8 kHz only, so an agent that needs 16 kHz telephony audio asks for PCM at 16000.
Bland Speech v3 ranked first of 13 AI voices in Design Arena's Audio Realism Bench results of 4 August 2026, at 1365, and the human recording scored 1500. Phone-agent prompts are one of the categories Design Arena uses. That result, mu-law at 8 kHz, and one standard rate of $0.04 per 1,000 characters are where Bland Speech fits for phone calls. A team sending short automated notices at the lowest price has Amazon Polly's Standard voices, at $4 per million characters, to weigh first.
Use Bland Speech when
- A team will send 8 kHz mu-law audio to a phone line and wants it from one request
- A developer is building on Pipecat or LiveKit and will add a phone leg through a carrier or SIP trunk
Use Amazon Polly when
- A team will send short automated notices on Polly's Standard voices at $4 per million characters
- A team already runs on AWS and will keep speech on its AWS bill
Pros
- 8 kHz mu-law and PCM from 8 to 48 kHz on the same endpoints
- A model that ranked first of 13 AI voices in Design Arena's results of 4 August 2026
- Streaming over HTTP and a WebSocket, with a service in Pipecat and a plugin in LiveKit
Cons
- More than a hobby project needs, because Bland Speech is built for production calls
The 8 kHz phone audio list orders APIs on this one question, the sample rates explainer covers what each rate changes, and the phone call list takes the wider view. Bland Speech has a studio for listening, the text to speech hub holds every comparison, and more explainers sit under text to speech guides.
Frequently asked questions
What is mu-law audio?
Mu-law is an 8-bit telephony encoding defined in G.711, which Google describes as 8-bit samples that compand 14-bit audio samples. Cartesia positions it for North American and Japanese telephony and for Twilio, and Bland offers it at 8 kHz only.
What sample rate do phone calls use?
Speechify says Twilio and LiveKit SIP expect 8 kHz mu-law or 16 kHz PCM, and Deepgram says requesting mu-law at 8 kHz matches Twilio's Media Streams format exactly. Bland's mu-law output is 8 kHz.
Is mu-law smaller than PCM?
Yes, at the same sample rate. Mono 8 kHz mu-law uses one byte per sample, 8,000 bytes a second, while mono 16-bit PCM at 8 kHz uses 16,000 bytes a second and at 48 kHz uses 96,000.
Why not request 48 kHz PCM for a phone call?
Speechify says Twilio and LiveKit SIP expect 8 kHz mu-law or 16 kHz PCM, so 48 kHz audio would need converting before it reaches those lines. Rime says to set a sampling rate of 8000 to match the telephony stream and shrink payloads, and Kalpa Labs names 48 kHz for WebRTC instead.
What does Bland Speech return by default?
A sample rate of 48000 Hz, the rate Bland Speech v3 renders natively. A request sets the encoding and rate, for example mulaw at 8000 Hz for a phone line.
