Audio Models
Text-to-Speech Models
Cartesia
| Model |
Price/1K chars |
Max chars |
Latency |
Languages |
Speed |
sonic-3 |
$0.065 |
50,000 |
90ms |
42 |
0.6-1.5x |
sonic-turbo |
$0.045 |
50,000 |
40ms |
15 |
0.6-1.5x |
sonic-2 |
$0.065 |
50,000 |
90ms |
15 |
0.6-1.5x |
sonic-3 is the recommended default — best balance of quality, speed, and language coverage. sonic-turbo is the fastest and cheapest option for latency-sensitive applications.
ElevenLabs
| Model |
Price/1K chars |
Max chars |
Latency |
Languages |
Speed |
eleven_flash_v2_5 |
$0.06 |
40,000 |
75ms |
32 |
0.7-1.2x |
eleven_turbo_v2_5 |
$0.06 |
40,000 |
250ms |
32 |
0.7-1.2x |
eleven_multilingual_v2 |
$0.12 |
40,000 |
300ms |
29 |
0.7-1.2x |
eleven_v3 |
$0.12 |
40,000 |
300ms |
74 |
0.7-1.2x |
ElevenLabs models support additional voice parameters: stability (0-1, default 0.5) and similarity_boost (0-1, default 0.75) for fine-grained voice control.
Output Formats
All TTS models support these output formats via the response_format parameter:
| Format |
Content-Type |
Description |
mp3 (default) |
audio/mpeg |
Compressed, smallest files |
wav |
audio/wav |
Uncompressed PCM in WAV container |
pcm |
audio/pcm |
Raw PCM 16kHz 16-bit |
pcm_16000 |
audio/pcm |
Raw PCM 16kHz 16-bit |
pcm_24000 |
audio/pcm |
Raw PCM 24kHz 16-bit |
ulaw_8000 |
audio/basic |
Mu-law 8kHz (telephony) |
mp3_22050_32 |
audio/mpeg |
MP3 22050Hz 32kbps |
mp3_44100_64 |
audio/mpeg |
MP3 44100Hz 64kbps |
mp3_44100_128 |
audio/mpeg |
MP3 44100Hz 128kbps |
Speech-to-Text Models (Batch)
Cartesia
| Model |
Price/min |
Max file |
Max duration |
Languages |
ink-whisper |
$0.00217 ($0.13/hr) |
10 MB |
10 min |
96 |
ink-whisper returns the transcript directly in the HTTP response — no polling required. Best for short-to-medium audio with fast turnaround.
AssemblyAI
| Model |
Price/min |
Max file |
Max duration |
Languages |
universal-3-pro |
$0.0035 |
10 MB |
4 hours |
99+ |
universal-2 |
$0.0025 |
10 MB |
4 hours |
99+ |
AssemblyAI models use an upload-then-poll workflow. The API blocks until transcription completes (typically 5-30 seconds). Best for longer audio and highest accuracy.
Supported Input Formats
| Provider |
Formats |
| Cartesia |
flac, m4a, mp3, mp4, mpeg, mpga, oga, ogg, wav, webm |
| AssemblyAI |
aac, flac, m4a, mp3, mp4, mpeg, mpga, mov, oga, ogg, wav, webm |
Response Formats
Control the output with the response_format parameter:
| Format |
Description |
json (default) |
{"text": "transcript"} |
text |
Plain text, no JSON wrapper |
verbose_json |
Includes text, language, duration, and words array with timestamps |
To get word-level timestamps, set response_format=verbose_json and add timestamp_granularities[]=word.
Speech-to-Text Models (Streaming)
Streaming models are used with the WebSocket endpoint. See Audio Streaming for the protocol guide.
AssemblyAI
| Model |
Price/min |
Billing |
Languages |
Max session |
u3-rt-pro |
$0.0075 |
Session time |
en, es, de, fr, pt, it |
3 hours |
universal-streaming-english |
$0.0025 |
Session time |
en |
3 hours |
universal-streaming-multilingual |
$0.0025 |
Session time |
en, es, de, fr, pt, it |
3 hours |
whisper-rt |
$0.005 |
Session time |
99+ |
3 hours |
Billing note: AssemblyAI streaming models bill by session wall-clock time. Silence and pauses count toward your bill.
Cartesia
| Model |
Price/min |
Billing |
Languages |
Max session |
ink-whisper |
$0.00433 ($0.26/hr) |
Audio time |
96 |
1 hour |
Billing note: Cartesia bills only for actual audio sent, not idle time.
Streaming Audio Requirements
| Provider |
Encodings |
Sample rates |
| AssemblyAI |
pcm_s16le, pcm_mulaw |
8000, 16000 Hz |
| Cartesia |
pcm_s16le, pcm_s32le, pcm_f16le, pcm_f32le, pcm_mulaw, pcm_alaw |
8000, 16000, 24000, 44100, 48000 Hz |
Streaming requires raw PCM audio — not MP3, WAV, or other container formats. Audio must be mono (single channel).