shunyalabs 3.0.3

Shunyalabs Python SDK — ASR, TTS, and Conversational AI

Shunyalabs Python SDK

The official Python SDK for Shunyalabs Speech AI APIs — ASR (speech-to-text) and TTS (text-to-speech).

Supports HTTP batch and WebSocket streaming modes with a fully async client.

Installation

pip install shunyalabs[all]

Install only what you need:

pip install shunyalabs[ASR]     # Speech-to-text only
pip install shunyalabs[TTS]     # Text-to-speech only
pip install shunyalabs[extras]  # Audio playback helpers (sounddevice)

Authentication

All API calls use Authorization: Bearer <api_key> header authentication.

from shunyalabs import AsyncShunyaClient

client = AsyncShunyaClient(api_key="your-api-key")

Or set the SHUNYALABS_API_KEY environment variable and omit api_key=.

---

Quick Start

TTS — Batch (HTTP)

import asyncio
from shunyalabs import AsyncShunyaClient
from shunyalabs.tts import TTSConfig

async def main():
    async with AsyncShunyaClient(api_key="your-api-key") as client:
        result = await client.tts.synthesize(
            "Hello, world!",
            config=TTSConfig(model="zero-indic", voice="Varun"),
        )
        result.save("output.mp3")
        print(f"{len(result.audio_data)} bytes saved")

asyncio.run(main())

TTS — Streaming (WebSocket)

import asyncio
from shunyalabs import AsyncShunyaClient
from shunyalabs.tts import TTSConfig

async def main():
    async with AsyncShunyaClient(api_key="your-api-key") as client:
        chunks = []
        async for audio in await client.tts.stream(
            "Hello, world!",
            config=TTSConfig(model="zero-indic", voice="Varun"),
        ):
            chunks.append(audio)
        print(f"{len(chunks)} chunks, {sum(len(c) for c in chunks)} bytes")

asyncio.run(main())

ASR — Batch (HTTP)

import asyncio
from shunyalabs import AsyncShunyaClient
from shunyalabs.asr import TranscriptionConfig

async def main():
    async with AsyncShunyaClient(api_key="your-api-key") as client:
        result = await client.asr.transcribe(
            "audio.wav",
            config=TranscriptionConfig(model="zero-indic"),
        )
        print(result.text)

asyncio.run(main())

ASR — Streaming (WebSocket)

import asyncio, subprocess
from shunyalabs import AsyncShunyaClient
from shunyalabs.asr import StreamingConfig, StreamingMessageType

async def main():
    async with AsyncShunyaClient(api_key="your-api-key") as client:
        conn = await client.asr.stream(
            config=StreamingConfig(language="en", sample_rate=16000),
        )

        @conn.on(StreamingMessageType.FINAL_SEGMENT)
        def on_seg(msg):
            print(f"[seg] {msg.text}")

        @conn.on(StreamingMessageType.FINAL)
        def on_final(msg):
            print(f"[final] {msg.text}")

        # Convert audio to 16kHz mono PCM and stream
        pcm = subprocess.run(
            ["ffmpeg", "-i", "audio.wav", "-ar", "16000", "-ac", "1", "-f", "s16le", "-"],
            capture_output=True,
        ).stdout

        for i in range(0, len(pcm), 4096):
            await conn.send_audio(pcm[i : i + 4096])

        await conn.end()
        await conn.close()

asyncio.run(main())

---

API Reference

Client Configuration

AsyncShunyaClient

Parameter	Type	Default	Description
api_key	str	None	API key. Falls back to SHUNYALABS_API_KEY env var.
timeout	float	60.0	Request timeout in seconds.
max_retries	int	2	Retries for failed requests (5xx, connection errors).
asr_url	str	https://asr.shunyalabs.ai	ASR batch API base URL.
asr_ws_url	str	wss://asr.shunyalabs.ai/ws	ASR streaming WebSocket URL.
tts_url	str	https://tts.shunyalabs.ai	TTS batch API base URL.
tts_ws_url	str	wss://tts.shunyalabs.ai/ws	TTS streaming WebSocket URL.

All URL parameters can also be set via environment variables: SHUNYALABS_ASR_URL, SHUNYALABS_ASR_WS_URL, SHUNYALABS_TTS_URL, SHUNYALABS_TTS_WS_URL.

Examples:

# Default — uses production endpoints
client = AsyncShunyaClient(api_key="your-api-key")

# Custom timeout and retries
client = AsyncShunyaClient(api_key="your-api-key", timeout=120.0, max_retries=5)

# Self-hosted endpoints
client = AsyncShunyaClient(
    api_key="your-api-key",
    asr_url="https://my-asr-server.example.com",
    tts_url="https://my-tts-server.example.com",
    tts_ws_url="wss://my-tts-server.example.com/ws",
)

---

TTS API

TTSConfig

Configuration for synthesis requests. Passed as config= to synthesize() and stream().

Parameter	Type	Default	Description
model	str	required	Model name (e.g. "zero-indic").
voice	str	required	Speaker voice name. See Available Speakers.
response_format	OutputFormat	"mp3"	Output audio format. See Output Formats.
speed	float	1.0	Speaking speed multiplier (0.25–4.0).
trim_silence	bool	False	Trim leading/trailing silence from audio.
volume_normalization	str	None	"peak" or "loudness".
background_audio	str	None	Preset name or base64-encoded background audio.
background_volume	float	0.1	Background volume relative to speech (0.0–1.0).

TTS Parameter Examples

model — Select the TTS model

# Currently available: "zero-indic"
config = TTSConfig(model="zero-indic", voice="Rajesh")
result = await client.tts.synthesize("Hello!", config=config)
# Output: 48000 bytes saved to output.mp3

voice — Choose a speaker

# Male English speaker
config = TTSConfig(model="zero-indic", voice="Varun")

# Female Hindi speaker
config = TTSConfig(model="zero-indic", voice="Sunita")

# Any speaker can speak any language — voice only controls vocal characteristics
config = TTSConfig(model="zero-indic", voice="Murugan")  # Tamil-native male speaking English
result = await client.tts.synthesize("Good morning, how are you?", config=config)

response_format — Output audio format

Values: "pcm", "wav", "mp3", "ogg_opus", "flac", "mulaw", "alaw"

# MP3 (default) — compressed, good for storage
config = TTSConfig(model="zero-indic", voice="Varun", response_format="mp3")
result = await client.tts.synthesize("Hello!", config=config)
result.save("output.mp3")
# Output: 12480 bytes (compressed)

# WAV — uncompressed, good for processing
config = TTSConfig(model="zero-indic", voice="Varun", response_format="wav")
result = await client.tts.synthesize("Hello!", config=config)
result.save("output.wav")
# Output: 96044 bytes (uncompressed with header)

# PCM — raw samples, for real-time pipelines
config = TTSConfig(model="zero-indic", voice="Varun", response_format="pcm")
result = await client.tts.synthesize("Hello!", config=config)
# Output: 96000 bytes (raw 16-bit samples)

# OGG Opus — compressed, good for web streaming
config = TTSConfig(model="zero-indic", voice="Varun", response_format="ogg_opus")

# mu-law / A-law — for telephony systems
config = TTSConfig(model="zero-indic", voice="Varun", response_format="mulaw")
config = TTSConfig(model="zero-indic", voice="Varun", response_format="alaw")

speed — Speaking speed multiplier

Range: 0.25 (very slow) to 4.0 (very fast). Default: 1.0.

# Slow — good for language learning
config = TTSConfig(model="zero-indic", voice="Nisha", speed=0.75)
result = await client.tts.synthesize("Take your time to understand this.", config=config)
# Output: longer audio, ~33% slower than normal

# Normal speed (default)
config = TTSConfig(model="zero-indic", voice="Nisha", speed=1.0)

# Fast — good for notifications or summaries
config = TTSConfig(model="zero-indic", voice="Nisha", speed=1.5)
result = await client.tts.synthesize("Quick update: your order has shipped.", config=config)
# Output: shorter audio, ~50% faster than normal

# Very fast
config = TTSConfig(model="zero-indic", voice="Nisha", speed=2.0)

trim_silence — Remove silence padding

# Without trim (default) — audio may have leading/trailing silence
config = TTSConfig(model="zero-indic", voice="Rajesh", trim_silence=False)
result = await client.tts.synthesize("Hello.", config=config)
# Output: 64000 bytes (includes silence padding)

# With trim — tighter audio, no dead air
config = TTSConfig(model="zero-indic", voice="Rajesh", trim_silence=True)
result = await client.tts.synthesize("Hello.", config=config)
# Output: 48000 bytes (silence stripped)

volume_normalization — Normalize audio loudness

Values: None (off), "peak", "loudness"

# No normalization (default)
config = TTSConfig(model="zero-indic", voice="Rajesh")

# Peak normalization — scale so the loudest sample hits 0 dBFS
config = TTSConfig(model="zero-indic", voice="Rajesh", volume_normalization="peak")
result = await client.tts.synthesize("This audio will have consistent peak levels.", config=config)

# Loudness normalization — perceptually even loudness (EBU R128)
config = TTSConfig(model="zero-indic", voice="Rajesh", volume_normalization="loudness")
result = await client.tts.synthesize("This audio will sound equally loud regardless of content.", config=config)

background_audio + background_volume — Add background music

import base64

# Using a preset name
config = TTSConfig(
    model="zero-indic",
    voice="Nisha",
    background_audio="cafe-ambience",
    background_volume=0.15,  # 15% volume relative to speech
)
result = await client.tts.synthesize("Welcome to our podcast.", config=config)

# Using custom audio (base64-encoded)
with open("background.mp3", "rb") as f:
    bg_b64 = base64.b64encode(f.read()).decode()

config = TTSConfig(
    model="zero-indic",
    voice="Nisha",
    background_audio=bg_b64,
    background_volume=0.1,  # 10% volume (subtle background)
)
result = await client.tts.synthesize("Welcome to our podcast.", config=config)
result.save("podcast_intro.mp3")

Available Speakers

Each speaker has a native language listed below, but every speaker can speak in any language — the native language only indicates the speaker's voice characteristics and accent.

Language	Male Speaker	Female Speaker
Assamese	Bimal	Anjana
Bengali	Arjun	Priyanka
Bodo	Daimalu	Hasina
Dogri	Vishal	Neelam
English	Varun	Nisha
Gujarati	Rakesh	Pooja
Hindi	Rajesh	Sunita
Kannada	Kiran	Shreya
Kashmiri	Farooq	Habba
Konkani	Mohan	Sarita
Maithili	Suresh	Meera
Malayalam	Krishnan	Deepa
Manipuri	Tomba	Ibemhal
Marathi	Siddharth	Ananya
Nepali	Bikash	Sapana
Odia	Bijay	Sujata
Punjabi	Gurpreet	Simran
Sanskrit	Vedant	Gayatri
Santali	Chandu	Roshni
Sindhi	Amjad	Kavita
Tamil	Murugan	Thangam
Telugu	Vishnu	Lakshmi
Urdu	Salman	Fatima

23 languages, 46 speakers (1 male + 1 female per language).

Expression Styles

Control the emotional tone by passing a style tag before the text (e.g. "<Happy> Hello!").

Style Tag	Description
<Happy>	Joyful, upbeat tone
<Sad>	Somber, melancholic tone
<Angry>	Forceful, intense tone
<Fearful>	Anxious, trembling tone
<Surprised>	Exclamatory, astonished tone
<Disgust>	Repulsed, disapproving tone
<News>	Formal news-anchor style
<Conversational>	Casual, everyday speech
<Narrative>	Storytelling / audiobook style
<Enthusiastic>	Energetic, passionate tone
<Neutral>	Clean read-speech (default, no tag needed)

Expression style examples:

# Happy greeting
config = TTSConfig(model="zero-indic", voice="Rajesh")
result = await client.tts.synthesize("<Happy> Welcome aboard! We're thrilled to have you.", config=config)

# News anchor reading
config = TTSConfig(model="zero-indic", voice="Nisha")
result = await client.tts.synthesize("<News> Breaking news: the markets rallied today.", config=config)

# Storytelling
config = TTSConfig(model="zero-indic", voice="Krishnan")
result = await client.tts.synthesize("<Narrative> Once upon a time, in a land far away...", config=config)

# Conversational chatbot
config = TTSConfig(model="zero-indic", voice="Simran")
result = await client.tts.synthesize("<Conversational> Hey! How's it going?", config=config)

# Neutral (default — no tag needed)
config = TTSConfig(model="zero-indic", voice="Varun")
result = await client.tts.synthesize("Your account balance is five thousand rupees.", config=config)

Output Formats

Format	Value
PCM (raw)	"pcm"
WAV	"wav"
MP3	"mp3"
OGG Opus	"ogg_opus"
FLAC	"flac"
mu-law	"mulaw"
A-law	"alaw"

TTS Methods

Batch (HTTP)

result = await client.tts.synthesize("text", config=TTSConfig(...))
result.save("output.mp3")       # Save to file
result.audio_data               # Raw bytes
result.sample_rate               # Sample rate (Hz)

Streaming (WebSocket)

# Iterate audio chunks
async for audio_bytes in await client.tts.stream("text", config=TTSConfig(...)):
    play(audio_bytes)

# With chunk metadata
async for chunk_meta, audio_bytes in await client.tts.stream("text", config=TTSConfig(...), detailed=True):
    print(chunk_meta.chunk_index, len(audio_bytes))

# Collect all and return combined bytes
audio = await client.tts.synthesize_stream("text", config=TTSConfig(...))

# Stream directly to file
await client.tts.stream_to_file("text", "output.pcm", config=TTSConfig(...))

TTSResult

Returned by synthesize().

Attribute	Type	Description
audio_data	bytes	Decoded audio bytes.
sample_rate	int	Audio sample rate in Hz.
format	str	Audio format string.

---

ASR API

Supported Audio & Video Formats

The following formats are natively supported. For video inputs, the audio track is automatically extracted before transcription begins.

Audio Formats

Format	Extension	Notes
WAV	.wav	Recommended — lossless, best accuracy
FLAC	.flac	Lossless compression
MP3	.mp3	Widely supported lossy format
M4A / AAC	.m4a, .aac	Common mobile/Apple format
OGG / OPUS	.ogg, .opus	Open container / low-latency codec
AMR / AMR-WB	.amr	Telephony standard format
AIFF	.aiff	Apple lossless audio format

Video Formats

Format	Extension
MP4	.mp4
WebM	.webm
MOV	.mov
MKV	.mkv

✓ For best transcription accuracy, use WAV files with PCM 16-bit encoding at 16,000 Hz mono.

Sample Rate Requirements

The API supports a range of sample rates to accommodate diverse audio sources including telephony recordings. Files provided at unlisted rates are automatically resampled during processing.

Sample Rate	Common Use Case
8,000 Hz	Telephony / phone call recordings
16,000 Hz	Standard speech recognition (recommended)
22,050 Hz	General audio
44,100 Hz	CD-quality audio
48,000 Hz	Professional broadcast audio

ℹ Audio provided at sample rates not listed above is automatically resampled to 16,000 Hz during processing. No manual conversion is required.

TranscriptionConfig

Configuration for batch transcription. Passed as config= to transcribe().

Models:

Model	Use for
zero-indic	General Indian languages (Hindi, Tamil, Telugu, Kannada, Marathi, Bengali, etc.)
zero-med	Medical/clinical audio — auto-applies medical terminology correction
zero-codeswitch	Code-switched speech (Hinglish, Tanglish, etc.) — auto-restores mixed text

Core Parameters:

Parameter	Type	Default	Description
model	str	required	Model name ("zero-indic", "zero-med", "zero-codeswitch").
language_code	str	"auto"	Language hint — ISO code (hi, ta, en) or name (Hindi, Tamil).
output_script	str	"auto"	Transliterate output script ("Latin", "Devanagari", "Tamil", etc.).
word_timestamps	bool	False	Per-word start/end times and alignment confidence scores.

Diarization & Speaker ID:

Parameter	Type	Default	Description
enable_diarization	bool	False	Identify who spoke when (SPEAKER_00, SPEAKER_01).
enable_speaker_identification	bool	False	Resolve speakers to registered names (requires diarization).
enable_emotion_diarization	bool	False	Detect dominant emotion per segment.
project	str	None	Speaker library namespace for speaker identification.

NLP Features:

Parameter	Type	Default	Description
enable_intent_detection	bool	False	Classify transcript intent via Gemini.
intent_choices	list[str]	None	Constrain to specific intents.
enable_summarization	bool	False	Generate concise summary via Gemini.
summary_max_length	int	150	Max words in summary.
enable_sentiment_analysis	bool	False	Sentiment label, score, and explanation.
enable_keyterm_normalization	bool	False	Normalize domain-specific terms.
keyterm_keywords	list[str]	None	Focus normalization on these terms.

Post-processing:

Parameter	Type	Default	Description
enable_profanity_hashing	bool	False	Mask profanity with ****.
hash_keywords	list[str]	None	Mask specific words/phrases with **** (regex).
output_language	str	None	Translate transcript to this language ("en", "hi", etc.).

ASR Parameter Examples

model + language_code — Basic transcription

# Auto-detect language (default)
config = TranscriptionConfig(model="zero-indic", language_code="auto")
result = await client.asr.transcribe("audio.wav", config=config)
print(result.text)
print(f"Detected: {result.detected_language}")
# Output:
#   "Hello, how are you doing today?"
#   Detected: English

# Specify language for better accuracy on short clips
config = TranscriptionConfig(model="zero-indic", language_code="hi")
result = await client.asr.transcribe("hindi_audio.wav", config=config)
print(result.text)
# Output: "नमस्ते, आप कैसे हैं?"

# Medical audio — auto-applies medical terminology correction
config = TranscriptionConfig(model="zero-med", language_code="en")

# Code-switched speech (Hinglish, Tanglish, etc.)
config = TranscriptionConfig(model="zero-codeswitch")

output_script — Transliterate output to a different script

Uses aksharamukha — no LLM involved.

Common values: "auto", "Latin", "Devanagari", "Bengali", "Telugu", "Tamil", "Kannada", "ITRANS"

# Latin/Roman script — romanised output
config = TranscriptionConfig(model="zero-indic", language_code="hi", output_script="Latin")
result = await client.asr.transcribe("hindi_audio.wav", config=config)
print(result.text)
# Output: "namaste mohammad ji ye ek zaruri call hai"

# Auto (default) — no transliteration
config = TranscriptionConfig(model="zero-indic", output_script="auto")

word_timestamps — Per-word timing and confidence

config = TranscriptionConfig(model="zero-indic", word_timestamps=True)
result = await client.asr.transcribe("audio.wav", config=config)
for seg in result.segments:
    if seg.words:
        for w in seg.words:
            print(f"  {w.word}  [{w.start:.3f}s - {w.end:.3f}s]  score={w.score}")
# Output:
#   नमस्ते  [0.532s - 0.932s]  score=-4.237
#   मोहम्मद  [1.012s - 1.412s]  score=-6.226

Note on score: Log-probability from the alignment model — more negative means lower confidence. Values above -5 are generally reliable.

enable_diarization — Who spoke when

config = TranscriptionConfig(model="zero-indic", enable_diarization=True)
result = await client.asr.transcribe("call.wav", config=config)
print(result.speakers)  # ["SPEAKER_00", "SPEAKER_01"]
for seg in result.segments:
    print(f"  [{seg.speaker}] [{seg.start:.1f}s - {seg.end:.1f}s] {seg.text}")
# Output:
#   [SPEAKER_00] [0.5s - 3.2s] नमस्ते, आप कैसे हैं
#   [SPEAKER_01] [4.1s - 6.8s] मैं ठीक हूँ धन्यवाद

enable_speaker_identification + project — Resolve to registered names

Requires enable_diarization=True and pre-registered voice profiles (see Speaker APIs).

config = TranscriptionConfig(
    model="zero-indic",
    enable_diarization=True,
    enable_speaker_identification=True,
    project="support_team",
)
result = await client.asr.transcribe("call.wav", config=config)
print(result.speakers)  # ["Priya", "Rahul"]
for seg in result.segments:
    print(f"  [{seg.speaker}] {seg.text}")
# Output:
#   [Priya] नमस्ते, आप कैसे हैं
#   [Rahul] मैं ठीक हूँ

enable_emotion_diarization — Emotion per segment

config = TranscriptionConfig(
    model="zero-indic",
    enable_diarization=True,
    enable_emotion_diarization=True,
)
result = await client.asr.transcribe("call.wav", config=config)
for seg in result.segments:
    print(f"  [{seg.speaker}] ({seg.emotion}) {seg.text}")
# Output:
#   [SPEAKER_00] (angry) I've been waiting for an hour!
#   [SPEAKER_01] (neutral) I'm sorry about that, let me help.

enable_intent_detection + intent_choices — Classify intent

# Constrained to specific choices
config = TranscriptionConfig(
    model="zero-indic",
    enable_intent_detection=True,
    intent_choices=["complaint", "inquiry", "service_request", "compliment"],
)
result = await client.asr.transcribe("customer_call.wav", config=config)
print(result.nlp_analysis.intent)
# Output: {"label": "service_request", "confidence": 0.92, "reasoning": "Caller is requesting roadside assistance"}

# Open intent — no constraints
config = TranscriptionConfig(model="zero-indic", enable_intent_detection=True)

enable_summarization + summary_max_length — Summarize transcript

config = TranscriptionConfig(
    model="zero-indic",
    enable_summarization=True,
    summary_max_length=50,  # max ~50 words
)
result = await client.asr.transcribe("meeting.wav", config=config)
print(result.nlp_analysis.summary)
# Output: "Customer called about a vehicle breakdown. Agent confirmed the complaint and promised a technician within the hour."

enable_sentiment_analysis — Detect sentiment

config = TranscriptionConfig(model="zero-indic", enable_sentiment_analysis=True)
result = await client.asr.transcribe("feedback.wav", config=config)
print(result.nlp_analysis.sentiment)
# Output: {"label": "negative", "score": -0.72, "explanation": "Customer expresses frustration about..."}

enable_keyterm_normalization + keyterm_keywords — Fix domain terms

config = TranscriptionConfig(
    model="zero-indic",
    enable_keyterm_normalization=True,
    keyterm_keywords=["EMI", "NACH mandate", "bounce charge"],
)
result = await client.asr.transcribe("finance_call.wav", config=config)
print(result.nlp_analysis.normalized_text)
# Output: "आपकी EMI की तारीख पाँच अगस्त है" (normalized from "emi")

enable_profanity_hashing + hash_keywords — Redact sensitive words

# Mask profanity
config = TranscriptionConfig(model="zero-indic", enable_profanity_hashing=True)
result = await client.asr.transcribe("audio.wav", config=config)
# Output: "अरे **** यह काम क्यों नहीं हो रहा"

# Mask specific keywords (regex, case-insensitive, no LLM)
config = TranscriptionConfig(
    model="zero-indic",
    hash_keywords=["account number", "card number", "OTP"],
)
result = await client.asr.transcribe("call.wav", config=config)
# Output: "आपका **** 4321 है और आपका **** कल भेजा गया था"

output_language — Translate transcript

config = TranscriptionConfig(
    model="zero-indic",
    language_code="hi",
    output_language="en",
)
result = await client.asr.transcribe("hindi_audio.wav", config=config)
print(result.text)
# Output: "Hello, this is an urgent call regarding your vehicle service."
print(result.nlp_analysis.translation)
# Also available in nlp_analysis

ASR Methods

Batch (HTTP)

The transcribe() method accepts multiple input types:

# From file path (string)
result = await client.asr.transcribe("audio.wav", config=TranscriptionConfig(model="zero-indic"))
result = await client.asr.transcribe("/absolute/path/to/recording.mp3", config=TranscriptionConfig(model="zero-indic"))

# From raw audio bytes (in-memory audio)
with open("audio.wav", "rb") as f:
    audio_bytes = f.read()
result = await client.asr.transcribe(audio_bytes, config=TranscriptionConfig(model="zero-indic"))

# From file object
with open("audio.wav", "rb") as f:
    result = await client.asr.transcribe_file(f, config=TranscriptionConfig(model="zero-indic"))

# From remote URL
result = await client.asr.transcribe_url("https://example.com/audio.wav", config=TranscriptionConfig(model="zero-indic"))

⚠ URL inputs must be publicly accessible at the time of the API request. Authenticated or signed URLs with short expiry times may fail.

Streaming (WebSocket)

conn = await client.asr.stream(config=StreamingConfig(language="en"))

StreamingConfig

Configuration for the WebSocket streaming session.

Parameter	Type	Default	Description
language	str	"auto"	Language code or "auto".
sample_rate	int	16000	Audio sample rate in Hz.
dtype	str	"int16"	Audio data type ("int16", "float32").
chunk_size_sec	float	1.0	Processing chunk size in seconds.
silence_threshold_sec	float	0.5	Silence duration to trigger segmentation.

Streaming Parameter Examples

language — Set recognition language

# Auto-detect (default)
conn = await client.asr.stream(config=StreamingConfig(language="auto"))

# Specific language for better accuracy
conn = await client.asr.stream(config=StreamingConfig(language="en"))
conn = await client.asr.stream(config=StreamingConfig(language="hi"))
conn = await client.asr.stream(config=StreamingConfig(language="ta"))

sample_rate + dtype — Match your audio source

# Standard microphone input: 16kHz, 16-bit integer (default)
conn = await client.asr.stream(config=StreamingConfig(
    sample_rate=16000,
    dtype="int16",
))

# High-quality audio: 48kHz, 32-bit float
conn = await client.asr.stream(config=StreamingConfig(
    sample_rate=48000,
    dtype="float32",
))

chunk_size_sec — Processing window size

# Smaller chunks = lower latency, more partial results
conn = await client.asr.stream(config=StreamingConfig(chunk_size_sec=0.5))

# Larger chunks = more context, potentially better accuracy
conn = await client.asr.stream(config=StreamingConfig(chunk_size_sec=2.0))

silence_threshold_sec — Control segment boundaries

# Quick segmentation — short pauses trigger a new segment
conn = await client.asr.stream(config=StreamingConfig(silence_threshold_sec=0.3))
# Good for: fast-paced dialogue, command recognition

# Patient segmentation — only split on longer pauses
conn = await client.asr.stream(config=StreamingConfig(silence_threshold_sec=1.5))
# Good for: lectures, monologues, dictation

Streaming Events

Register event handlers on the connection object:

conn = await client.asr.stream(config=StreamingConfig(language="en"))

@conn.on(StreamingMessageType.PARTIAL)
def on_partial(msg):
    print(f"Interim: {msg.text}")

@conn.on(StreamingMessageType.FINAL_SEGMENT)
def on_segment(msg):
    print(f"Segment: {msg.text}")

@conn.on(StreamingMessageType.FINAL)
def on_final(msg):
    print(f"Final: {msg.text} ({msg.audio_duration_sec}s)")

@conn.on(StreamingMessageType.DONE)
def on_done(msg):
    print(f"Done. {msg.total_segments} segments, {msg.total_audio_duration_sec}s")

@conn.on(StreamingMessageType.ERROR)
def on_error(msg):
    print(f"Error: {msg.message}")

Event	Model	Key Attributes
PARTIAL	StreamingPartial	text, language, segment_id, latency_ms
FINAL_SEGMENT	StreamingFinalSegment	text, language, segment_id, silence_duration_ms
FINAL	StreamingFinal	text, language, audio_duration_sec, inference_time_ms
DONE	StreamingDone	total_segments, total_audio_duration_sec
ERROR	StreamingError	message, code

Streaming Connection Methods

await conn.send_audio(pcm_bytes)   # Send raw PCM audio
await conn.end()                    # Signal end of audio stream
await conn.close()                  # Close WebSocket connection
conn.is_closed                      # Check connection status
conn.session_id                     # Server-assigned session ID

TranscriptionResult

Returned by transcribe().

Response Schema

{
  "success": true,
  "request_id": "a1b2c3...",
  "text": "[Priya] नमस्ते मेरी गाड़ी खराब हो गई है [Rahul] ठीक है मैं आपकी मदद करता हूँ",
  "segments": [
    {
      "start": 0.5,
      "end": 4.2,
      "text": "नमस्ते मेरी गाड़ी खराब हो गई है",
      "speaker": "Priya",
      "emotion": "sad",
      "words": [
        { "word": "नमस्ते", "start": 0.53, "end": 0.93, "score": -4.2 }
      ]
    }
  ],
  "detected_language": "hindi",
  "speakers": ["Priya", "Rahul"],
  "audio_duration": 9.1,
  "inference_time_ms": 3241.5,
  "nlp_analysis": {
    "intent": { "label": "service_request", "confidence": 0.95, "reasoning": "..." },
    "summary": "Customer reported a vehicle breakdown. Agent offered assistance.",
    "sentiment": { "label": "negative", "score": -0.6, "explanation": "..." },
    "normalized_text": "...",
    "translation": "..."
  }
}

Accessing Fields

result = await client.asr.transcribe("audio.wav", config=TranscriptionConfig(model="zero-indic"))

# Full transcript
print(result.text)

# Detected language
print(result.detected_language)

# Speakers (when diarization is enabled)
print(result.speakers)  # ["SPEAKER_00", "SPEAKER_01"] or ["Priya", "Rahul"]

# Iterate timestamped segments
for seg in result.segments:
    speaker = f"[{seg.speaker}] " if seg.speaker else ""
    emotion = f"({seg.emotion}) " if seg.emotion else ""
    print(f"{speaker}{emotion}[{seg.start:.2f}s → {seg.end:.2f}s]  {seg.text}")

    # Per-word timestamps (when word_timestamps=True)
    if seg.words:
        for w in seg.words:
            print(f"    {w.word} [{w.start:.3f}s-{w.end:.3f}s] score={w.score}")

Attribute	Type	Description
success	bool	Whether transcription succeeded.
request_id	str	Unique request identifier.
text	str	Full transcription text (prefixed with [SPEAKER_XX] when diarized).
segments	list[SegmentResult]	Time-aligned segments (see below).
detected_language	str	Detected language name (e.g. "hindi", "English").
speakers	list[str]	Unique speaker labels (empty when diarization is off).
audio_duration	float	Audio duration in seconds.
inference_time_ms	float	Server inference time in ms.
nlp_analysis	NLPAnalysis	NLP results (if any enable_* flags were set).

SegmentResult

Attribute	Type	Description
start	float	Segment start time in seconds.
end	float	Segment end time in seconds.
text	str	Transcribed text for this segment.
speaker	str or None	Speaker label (when diarization is enabled).
emotion	str or None	Emotion label (when emotion diarization is enabled).
words	list[WordResult] or None	Per-word timestamps (when word_timestamps=True).

WordResult

Attribute	Type	Description
word	str	The word.
start	float	Word start time in seconds.
end	float	Word end time in seconds.
score	float or None	Log-probability confidence (more negative = less confident).

NLPAnalysis

Attribute	Type	Description
intent	dict	Intent label, confidence, reasoning.
summary	str	Transcript summary.
sentiment	dict	Sentiment label, score, explanation.
emotion	dict	Emotion detection results.
translation	str	Translated text (when output_language is set).
normalized_text	str	Text with normalized key terms.

---

Exceptions

All exceptions inherit from ShunyalabsError.

Exception	Description
AuthenticationError	Invalid or missing API key (401).
PermissionDeniedError	Insufficient permissions (403).
NotFoundError	Resource not found (404).
RateLimitError	Rate limit exceeded (429).
ServerError	Server-side error (5xx).
TimeoutError	Request timed out.
ConnectionError	Network connectivity issue.
ConfigurationError	Invalid or missing configuration.
AudioError	Invalid, corrupted, or unsupported audio data.
SessionError	Invalid or expired session state.
ConversationError	Error returned during conversation flow.
TranscriptionError	ASR-specific transcription error.
SynthesisError	TTS-specific synthesis error.

Exception handling example:

from shunyalabs import (
    ConfigurationError,
    AuthenticationError,
    ConnectionError,
    TranscriptionError,
    AudioError,
    SessionError,
    TimeoutError,
    ConversationError,
)

try:
    result = await client.asr.transcribe("audio.wav", config=TranscriptionConfig(model="zero-indic"))
    print(result.text)

except ConfigurationError as e:
    print(f"Invalid configuration: {e}")

except AuthenticationError as e:
    print(f"Authentication failed: {e}")

except ConnectionError as e:
    print(f"Unable to connect: {e}")

except TranscriptionError as e:
    print(f"Transcription processing failed: {e}")

except AudioError as e:
    print(f"Invalid/unsupported audio: {e}")

except SessionError as e:
    print(f"Invalid or expired session: {e}")

except TimeoutError as e:
    print(f"Request timed out: {e}")

except ConversationError as e:
    print(f"Conversation error: {e}")

except Exception as e:
    print(f"Unexpected error: {e}")

---

---

File Duration & Rate Limits

Limit	Value
Maximum audio duration	15 minutes
Maximum File size	50 MB
Concurrent requests	16

⚠ For files exceeding the duration limit, split the audio into smaller segments before submitting. Consider using silence-detection tools to identify natural break points.

ℹ Support for pre-signed URLs for large audio files will be introduced in an upcoming beta release, allowing clients to provide a secure URL instead of uploading directly.

---

ASR End-to-End Quickstart

A complete integration walkthrough from installation to parsing the response.

Step 1 — Install the SDK

pip install shunyalabs[all]

Step 2 — Authenticate

import os
from shunyalabs import AsyncShunyaClient

client = AsyncShunyaClient(api_key=os.environ.get("SHUNYALABS_API_KEY"))

⚠ Store your API key in an environment variable rather than hardcoding it.

Step 3 — Transcribe Audio

from shunyalabs.asr import TranscriptionConfig

async with AsyncShunyaClient(api_key=os.environ.get("SHUNYALABS_API_KEY")) as client:
    result = await client.asr.transcribe(
        "recording.wav",
        config=TranscriptionConfig(model="zero-indic"),
    )

Step 4 — Parse & Use the Response

# Full transcript text
print("Transcript:", result.text)

# Detected language
print("Language:", result.detected_language)

# Timestamped segments
for seg in result.segments:
    print(f"  [{seg.start:.2f}s – {seg.end:.2f}s]  {seg.text}")

Complete Example

import asyncio, os
from shunyalabs import AsyncShunyaClient, AuthenticationError
from shunyalabs.asr import TranscriptionConfig

async def main():
    async with AsyncShunyaClient(api_key=os.environ.get("SHUNYALABS_API_KEY")) as client:
        config = TranscriptionConfig(
            model="zero-indic",
            enable_diarization=True,
            word_timestamps=True,
            enable_sentiment_analysis=True,
        )
        try:
            result = await client.asr.transcribe("customer_call.wav", config=config)

            print("=== Transcript ===")
            print(result.text)
            print(f"Detected language: {result.detected_language}")
            print(f"Speakers: {result.speakers}")
            print()
            print("=== Segments ===")
            for seg in result.segments:
                speaker = f"[{seg.speaker}] " if seg.speaker else ""
                print(f"{speaker}[{seg.start:.2f}s → {seg.end:.2f}s]  {seg.text}")

            if result.nlp_analysis and result.nlp_analysis.sentiment:
                print(f"\nSentiment: {result.nlp_analysis.sentiment}")

        except AuthenticationError as e:
            print(f"Authentication failed: {e}")
        except Exception as e:
            print(f"Error: {e}")

asyncio.run(main())

---

Development

git clone https://github.com/Shunyalabsai/shunyalabs-python-sdk.git
cd shunyalabs-python-sdk

python -m venv .venv
source .venv/bin/activate
pip install -e ".[dev]"

# Run tests
pytest

# Lint
ruff check src/
black --check src/
mypy src/

License

MIT