Text-to-Speech

Streaming Text To Speech

SDK

Use the SDK for safer, faster integration than manual HTTP calls.

Generate speech from text using real-time streaming, allowing audio playback to begin before the entire synthesis is complete.

This endpoint streams audio data in chunks, enabling low-latency audio playback for applications requiring immediate feedback.

Streaming Format:

WAV format: First chunk contains WAV header (size=0xFFFFFFFF for streaming) followed by raw PCM data. Subsequent chunks contain only PCM data.
MP3 format: Each chunk contains post-processed MP3 data that can be decoded independently.

Use Cases:

Conversational AI, chatbots and real-time voice assistants
Interactive applications requiring immediate audio feedback
Long-form content where waiting for full synthesis is impractical

Request Parameters: Uses the same TTSRequest schema as the standard TTS endpoint. Set output.audio_format to "wav" or "mp3" to control the streaming format.

POST/v1/text-to-speech/stream

cURL

# Pipe streaming audio directly into ffplay for real-time playback.
# Requires: ffmpeg (brew/choco/apt install ffmpeg)
curl -N -s --request POST \
  --url https://api.typecast.ai/v1/text-to-speech/stream \
  --header 'Content-Type: application/json' \
  --header 'X-API-KEY: <api-key>' \
  --data @- <<EOF | ffplay -autoexit -nodisp -loglevel error -i pipe:0
{
  "voice_id": "tc_60e5426de8b95f1d3000d7b5",
  "text": "Thanks for reaching out. Your reservation has been confirmed for Friday at 7 PM.",
  "model": "ssfm-v30"
}
EOF

"[Binary audio stream - WAV chunks]"

Authorizations

X-API-KEYstringheaderrequired

API key for authentication. You can obtain an API key from the Typecast API Console.

Body

application/json

Text-to-speech streaming request parameters

seedinteger

Unsigned integer seed for reproducible speech generation. The same seed with the same input parameters will produce identical audio output.

Must be a non-negative integer (≥ 0). Negative values are not accepted.
If omitted, the server generates a random seed each time, producing slight variations.

textstringrequired

Text to convert to speech. Minimum 1 character, maximum 2000 characters. Credits consumed based on text length. Supports multiple languages including English, Korean, Japanese, and Chinese. Special characters and punctuation are handled automatically.

modelssfm-v30 | ssfm-v21required

Voice model to use for speech synthesis.

ssfm-v30: Latest model with improved prosody and additional emotion presets (recommended)
ssfm-v21: Stable production model with reliable quality

ssfm-v30ssfm-v21

outputobject

Streaming audio output settings including pitch (-12 to +12 semitones), tempo (0.5x to 2.0x), format (wav/mp3), and target_lufs (-70 to 0 LUFS). Note: volume is not available in streaming mode.

Also supports remove_silence_ms (integer, 0–1000 ms) for silence removal.

Show nested model

audio_pitchinteger

Adjusts the pitch in semitones to affect perceived gender and age: -12 (one octave lower, deeper voice), -6 (half octave lower), 0 (original pitch, default), +6 (half octave higher), +12 (one octave higher, higher voice)

audio_temponumber

Controls speech speed: 0.5 (half speed, very slow and clear), 0.75 (slightly slower than normal), 1.0 (normal speaking speed, default), 1.5 (50% faster than normal), 2.0 (double speed, very fast speech)

target_lufsnumber

Sets the target absolute loudness (LUFS) for streaming output audio. This normalizes generated audio to a consistent loudness regardless of the original source. Cannot be used with the volume parameter.

Recommended values: -14 (common streaming standard), -23 (broadcast standard).

audio_formatwav | mp3

Output audio format for streaming.

WAV format:

Uncompressed PCM audio
16-bit depth, mono channel, 32000 Hz sample rate
Chunked transfer: first chunk contains the WAV header (size = 0xFFFFFFFF), subsequent chunks contain raw PCM data
Recommended when you want to play audio as it arrives

MP3 format:

Compressed MPEG Layer III audio
320 kbps bitrate, 44100 Hz sample rate
Chunked transfer: each chunk contains independently decodable MPEG frames
Recommended for bandwidth-constrained clients

wavmp3

remove_silence_msinteger

When enabled, shortens detected silences longer than the specified duration to that duration. The value is in milliseconds (ms). The value is the duration of silence to retain, not the amount to remove.

Accepted values:

An integer from 0 to 1000, with a recommended range of 0 to 200.
Omitted or null: silence removal is disabled.
0: removes detected silence longer than 0 ms. This does not disable the feature.
Booleans, strings, fractional values, and out-of-range values are invalid.

Example: 100 shortens detected silences longer than 100 ms to 100 ms. Lower values include shorter silence segments for removal and leave less silence in each affected segment.

promptPrompt

Emotion and style settings for the generated speech, including emotion type (happy/sad/angry/normal) and intensity (0.0 to 2.0) to control the emotional expression

SmartPrompt (ssfm-v30)

next_textstring

Text that comes AFTER the text field in TTSRequest. Provides forward context for emotion inference.

The model analyzes the flow: previous_text → text (synthesized) → next_text

Maximum 2000 characters
Helps the model anticipate emotional transitions
Leave empty if no following context is available

emotion_typestring

Discriminator field to identify the prompt type. Must be set to "smart" for context-aware emotion inference.

previous_textstring

Text that comes BEFORE the text field in TTSRequest. Provides backward context for emotion inference.

The model analyzes the flow: previous_text → text (synthesized) → next_text

Maximum 2000 characters
Helps the model understand emotional build-up and context
Leave empty if no preceding context is available

PresetPrompt (ssfm-v30)

emotion_typestring

Discriminator field to identify the prompt type. Must be set to "preset" for preset-based emotion control.

Emotion preset to apply to the generated speech.

Supported emotions: normal, happy, sad, angry, whisper, toneup, tonedown

Check available emotions for each voice through the /v2/voices API.

normalsadhappyangrywhispertoneuptonedown

emotion_intensitynumber

Controls the strength of emotional expression in the generated speech.

0.0: Completely neutral, no emotional coloring
0.5: Subtle emotional hints
1.0: Standard emotional expression (default)
1.5: Strong emotional emphasis
2.0: Maximum intensity, highly expressive

Prompt (ssfm-v21)

emotion_presetobject

Emotion preset to apply.

Supported emotions for ssfm-v21: normal, happy, sad, angry

Check available emotions for each voice through the /v2/voices API.

emotion_intensityobject

Controls the strength of emotional expression (0.0 to 2.0).

0.0: Completely neutral
1.0: Standard expression (default)
2.0: Maximum intensity

languagestring

Language code following ISO 639-3 standard. Case-insensitive (both "ENG" and "eng" are accepted). If not provided, will be auto-detected based on text content.

ssfm-v30 Supported Languages (37)

Code	Language
ARA	Arabic
IND	Indonesian
POR	Portuguese
BEN	Bengali
ITA	Italian
RON	Romanian
BUL	Bulgarian
JPN	Japanese
RUS	Russian
CES	Czech
KOR	Korean
SLK	Slovak
DAN	Danish
MSA	Malay
SPA	Spanish
DEU	German
NAN	Min Nan
SWE	Swedish
ELL	Greek
NLD	Dutch
TAM	Tamil
ENG	English
NOR	Norwegian
TGL	Tagalog
FIN	Finnish
PAN	Punjabi
THA	Thai
FRA	French
POL	Polish
TUR	Turkish
HIN	Hindi
UKR	Ukrainian
VIE	Vietnamese
HRV	Croatian
YUE	Cantonese
ZHO	Chinese
HUN	Hungarian

ssfm-v21 Supported Languages (27)

Code	Language
ARA	Arabic
IND	Indonesian
RON	Romanian
BUL	Bulgarian
ITA	Italian
RUS	Russian
CES	Czech
JPN	Japanese
SLK	Slovak
DAN	Danish
KOR	Korean
SPA	Spanish
DEU	German
MSA	Malay
SWE	Swedish
ELL	Greek
NLD	Dutch
TAM	Tamil
ENG	English
POL	Polish
TGL	Tagalog
FIN	Finnish
POR	Portuguese
UKR	Ukrainian
FRA	French
HRV	Croatian
ZHO	Chinese

voice_idstringrequired

Voice identifier. Two prefixes are supported:

tc_ — Built-in Typecast voices (e.g., tc_60e5426de8b95f1d3000d7b5). See Listing all voices for available IDs.
uc_ — Custom voices created via Instant cloning (e.g., uc_64a1b2c3d4e5f6a7b8c9d0e1). Only the owner of a cloned voice can use it.

Case-sensitive: must use lowercase prefix.

Response

200Success - Returns streaming audio data in chunksaudio/wav

cURL

# Pipe streaming audio directly into ffplay for real-time playback.
# Requires: ffmpeg (brew/choco/apt install ffmpeg)
curl -N -s --request POST \
  --url https://api.typecast.ai/v1/text-to-speech/stream \
  --header 'Content-Type: application/json' \
  --header 'X-API-KEY: <api-key>' \
  --data @- <<EOF | ffplay -autoexit -nodisp -loglevel error -i pipe:0
{
  "voice_id": "tc_60e5426de8b95f1d3000d7b5",
  "text": "Thanks for reaching out. Your reservation has been confirmed for Friday at 7 PM.",
  "model": "ssfm-v30"
}
EOF

"[Binary audio stream - WAV chunks]"