Tools & Products

Sarvam AI Launches Saaras V4: 3B Hybrid State-Space Speech Model, 22 Indian Languages, Single-Pass Diarization, and Telephony Benchmarks

Sarvam AI released Saaras V4, a speech-to-text model combining a neural audio encoder with a 3B hybrid state-space language model. Saaras V4 records a 2.9% language identification error rate, provides single-pass diarization across 22 Indian languages and Global English, and processes 8kHz telephony audio at $0.30 per hour.

By FreakVinci · 2026-09-23 · 17 min read

On September 23, 2026, Bengaluru-based AI research laboratory Sarvam AI introduced Saaras V4, its fourth-generation speech recognition foundation model. Engineered specifically for complex acoustic environments and high-variance multilingual speech, Saaras V4 addresses the persistent failure modes of Western speech engines on Indian phonetics, accents, and code-mixed vernaculars.

Rather than fine-tuning a generic open-weights transformer, Sarvam AI architected Saaras V4 around a custom 3-billion-parameter hybrid state-space language decoder paired with a multi-band acoustic encoder. The model achieves state-of-the-art Word Error Rate (WER) across all 22 official Indian languages and Global English, offers native single-pass speaker diarization, and processes 8kHz telephony audio at $0.30 per audio hour.

1. Specifications & Model Card

Specification Saaras V4 (Sarvam AI) OpenAI Whisper large-v3 Google USM / Chirp
Developer Sarvam AI (Bengaluru, India) OpenAI Google Cloud
Model Type Audio Encoder + 3B Hybrid SSM Decoder Encoder-Decoder Transformer Conformer-CTC / Transducer
Supported Languages 22 Indian Languages + Global & Indian English 99 Languages (Broad global) 100+ Languages
Language Identification (LID) Error 2.9% (Top 10 Indian) / 5.22% (All 22) ~14.8% on regional Indian dialects ~7.1% on Indic dialects
Code-Mixed Speech Handling Native (Hinglish, Tanglish, etc.) Frequent phonetic hallucination Moderate
Integrated Diarization Single-Pass (No external pipeline) External model required (PyAnnote) External pipeline required
Telephony Acoustic Adaptation Dedicated 8kHz & 16kHz Co-Training General wideband (16kHz focused) General wideband
Output Representations 5 Modes (Verbatim, Normalized, etc.) Single mode (Normalized text) Normalized text
API Pricing $0.30 per audio hour ($0.005/min) $0.36 per hour ($0.006/min) $0.96 per hour ($0.016/min)
API Model Identifier saaras:v4 whisper-1 chirp-2

2. Neural Architecture: Hybrid State-Space Decoder

Standard automatic speech recognition models, including Whisper, rely on standard multi-head self-attention across both the audio encoder and the autoregressive text decoder. While attention captures global context, its quadratic $O(N^2)$ sequence scaling creates a computational bottleneck during long-form audio transcription, especially when maintaining deep token history for speaker tracking.

Sarvam AI engineered a hybrid architecture combining a continuous convolutional-transformer audio encoder with a 3B parameter hybrid state-space model (SSM):

                               [Raw Audio Input: 8kHz / 16kHz]
                                              │
                                              ▼
                             ┌──────────────────────────────────┐
                             │ Multi-Resolution Audio Front-End │
                             │  - Log-Mel Filterbank Extraction │
                             │  - Spectral Noise Masking        │
                             └────────────────┬─────────────────┘
                                              │
                                              ▼
                             ┌──────────────────────────────────┐
                             │   Conformer Audio Encoder        │
                             │  (Local Depthwise Convolutions + │
                             │   Relative Position Attention)   │
                             └────────────────┬─────────────────┘
                                              │ Continuous Audio Embeddings
                                              ▼
                             ┌──────────────────────────────────┐
                             │ Cross-Attention Projection Layer │
                             └────────────────┬─────────────────┘
                                              │
                                              ▼
                             ┌──────────────────────────────────┐
                             │ 3B Hybrid State-Space Decoder    │
                             │  - Linear-Time SSM Recurrence    │
                             │  - Interleaved Gated Attention   │
                             │  - Single-Pass Speaker Tracker   │
                             └────────────────┬─────────────────┘
                                              │
                    ┌─────────────────────────┴─────────────────────────┐
                    ▼                                                   ▼
       [Token Generation Stream]                          [Speaker Turn Attribution]
     - Verbatim / Normalized Tokens                     - Speaker ID Labels (Spk 0, Spk 1)

A. Linear-Time Sequence Processing

By replacing standard dense attention in the decoder with hybrid state-space layers (similar to Mamba-2 blocks interleaved with gated linear projections), Saaras V4 evaluates long context windows in linear $O(N)$ time. The decoder maintains a continuous memory state representing preceding conversation turns without allocating large Key-Value (KV) cache buffers.

B. Single-Pass Diarization Engine

Traditional transcription setups run two separate models: an ASR model to decode words, and an acoustic clustering model (such as PyAnnote) to estimate speaker embeddings and segment timestamps. This multi-stage process doubles inference latency and frequently misassigns words spoken during rapid dialogue.

Saaras V4 predicts speaker tokens directly within the target vocabulary stream. The language decoder learns joint acoustic and lexical cues, attributing speaker switches (e.g., <speaker:1> to <speaker:2>) synchronously with text generation.

3. Five Configurable Output Representation Modes

Speech applications require different formatting depending on whether the output feeds a legal transcript, an LLM agent, or a customer-facing screen. Saaras V4 natively provides five distinct text formats directly from the model, eliminating post-processing scripts:

Five Native Representation Modes in Saaras V4

Input Audio: "arre suno kal subah 9 baje meeting hai okay?"

1. Verbatim Mode:
   "arre suno kal subah 9 baje meeting hai okay?"

2. Normalized Mode:
   "Arre suno, kal subah 9:00 AM meeting hai, okay?"

3. Code-Mixed Mode:
   "अरे सुनो, कल सुबह 9:00 AM meeting है, okay?"

4. Transliteration Mode (Roman to Devanagari or inverse):
   "अरे सुनो कल सुबह ९ बजे मीटिंग है ओके?"

5. Direct Translation Mode (to English):
   "Listen, there is a meeting tomorrow morning at 9:00 AM, okay?"
  1. Verbatim Transcription: Captures every acoustic syllable, including false starts, repetitions, and vocalized pauses (umm, uh-huh, arre). Essential for psychiatric evaluations, legal depositions, and speech pathology research.
  2. Normalized Text: Standardizes spoken numbers into Arabic numerals (forty five thousand becomes 45,000), applies standard punctuation, and fixes capitalization for readable display.
  3. Code-Mixed Text: Preserves intra-sentence language switching (such as Hinglish, Tanglish, or Benglish) in its natural phonemic representation without forcing an arbitrary monolingual dictionary.
  4. Transliteration: Automatically converts native script utterances into phonetic Latin characters, or maps spoken English loanwords into local Indic scripts.
  5. Direct Speech Translation: Generates fluent English text directly from non-English audio inputs in a single decoding pass.

4. Benchmark Evaluations: Word Error Rate & Code-Mixing

Sarvam AI evaluated Saaras V4 across standardized public benchmarks and real-world audio datasets, focusing on noisy environments, regional accents, and financial call centers.

A. Word Error Rate (WER) across Indic Languages

Lower scores indicate fewer insertion, deletion, and substitution errors per 100 spoken words.

Language Saaras V4 WER (%) Whisper large-v3 WER (%) Google USM / Chirp WER (%)
Hindi (Standard) 6.4% 11.2% 8.9%
Hindi (Bihari / UP Accents) 8.1% 18.7% 13.4%
Tamil 9.2% 16.4% 12.1%
Telugu 8.8% 15.9% 11.7%
Bengali 7.6% 14.1% 10.3%
Marathi 8.3% 17.2% 12.8%
Kannada 9.7% 19.5% 14.2%
Gujarati 7.9% 15.3% 11.5%
Punjabi 8.4% 16.8% 13.0%
Malayalam 10.1% 21.4% 15.6%

B. Global English Benchmark Evaluations

To confirm that Indic specialization did not compromise English performance, Sarvam AI benchmarked Saaras V4 across seven recognized global English benchmarks:

English Benchmark Comparison (Word Error Rate - Lower is Better)

LibriSpeech Test-Clean
Saaras V4             ██ 1.8%
Whisper large-v3      ██ 1.9%
Google USM            ██ 2.1%

LibriSpeech Test-Other (Acoustic Noise)
Saaras V4             ████ 3.6%
Whisper large-v3      █████ 4.2%
Google USM            ████ 3.9%

Earnings-21 (Corporate Financial Earnings Calls)
Saaras V4             ██████ 5.8%
Whisper large-v3      █████████ 8.4%
Google USM            ███████ 6.9%

CHiME-7 (Severe Multi-Speaker Background Noise)
Saaras V4             ██████████ 9.8%
Whisper large-v3      ███████████████ 14.6%
Google USM            ████████████ 11.9%

C. Language Identification (LID) Precision

On multi-speaker dialogues where participants alternate between English and regional languages mid-sentence, Saaras V4 achieved a 2.9% LID error rate on the top 10 Indian languages and 5.22% across all 22 official languages, compared to 14.8% for Whisper large-v3.

5. Telephony & Acoustic Robustness

Most consumer speech recognition benchmarks evaluate 16kHz, high-fidelity uncompressed studio recordings. In commercial deployments (such as banking customer service lines, logistics dispatch, and telecom hotlines), audio arrives over 8kHz band-limited cellular or PSTN channels with background traffic noise.

Acoustic Degradation Spectrum Handled by Saaras V4:

[8kHz PSTN / GSM Adaptive Multi-Rate Codec] ──► [Packet Loss Concealment Jitter]
                                                              │
                                                              ▼
                                               [Ambient Street / Call Center Noise]
                                                              │
                                                              ▼
                                               [Saaras V4 Spectral De-noising Layer]
                                                              │
                                                              ▼
                                               [Accurate Word & Entity Extraction]

Technical Mitigations in Saaras V4

  1. Multi-Rate Acoustic Training: Rather than relying on simple upsampling, Sarvam AI trained the convolutional audio front-end on mixed batches containing native 8kHz telephony recordings alongside 16kHz and 48kHz audio.
  2. Noise-Augmented Spectrograms: The training regimen incorporated dynamic SpecAugment masking overlaid with acoustic profiles recorded in Indian railway stations, roadside auto-rickshaws, and dense BPO call centers.
  3. Financial and Legal Entity Retention: Indian earnings calls and banking hotlines require precise recognition of company names, stock tickers, and numerical quantities (e.g., lakhs and crores). Saaras V4 preserves Indian numbering syntax without mistranscribing numerical orders of magnitude.

6. Developer Implementation: Sarvam API Integration

Developers can deploy Saaras V4 through Sarvam AI’s cloud API. The model endpoint handles synchronous transcription and batched audio files.

A. Python Client Implementation

import requests

API_URL = "https://api.sarvam.ai/speech-to-text"
API_KEY = "YOUR_SARVAM_API_KEY"

headers = {
    "api-subscription-key": API_KEY
}

# Payload configuring Saaras V4 with single-pass diarization
payload = {
    "model": "saaras:v4",
    "language_code": "hi-IN",          # Target primary language (or auto-detect)
    "with_diarization": "true",        # Enable single-pass speaker identification
    "output_mode": "normalized",       # Choose from verbatim, normalized, code-mixed
    "with_timestamps": "true"          # Return word-level start and end timestamps
}

files = {
    "file": ("call_recording_8khz.wav", open("call_recording_8khz.wav", "rb"), "audio/wav")
}

response = requests.post(API_URL, headers=headers, data=payload, files=files)
result = response.json()

print(f"Transcript: {result['transcript']}")
for turn in result.get("diarization", []):
    print(f"[{turn['speaker']}] ({turn['start_time']}s - {turn['end_time']}s): {turn['text']}")

B. Direct cURL Command

curl -X POST "https://api.sarvam.ai/speech-to-text" \
  -H "api-subscription-key: $SARVAM_API_KEY" \
  -F "file=@customer_support_call.mp3" \
  -F "model=saaras:v4" \
  -F "language_code=unknown" \
  -F "output_mode=code-mixed" \
  -F "with_diarization=true"

C. Sample JSON Response with Single-Pass Diarization

{
  "request_id": "req_srvm_994829104",
  "language_code": "hi-IN",
  "confidence": 0.982,
  "transcript": "नमस्ते, मैं आपकी क्या सहायता कर सकता हूँ? Yes, I need help checking my account balance.",
  "diarization": [
    {
      "speaker": "Speaker_0",
      "start_time": 0.12,
      "end_time": 2.45,
      "text": "नमस्ते, मैं आपकी क्या सहायता कर सकता हूँ?"
    },
    {
      "speaker": "Speaker_1",
      "start_time": 2.68,
      "end_time": 5.10,
      "text": "Yes, I need help checking my account balance."
    }
  ]
}

7. Economic Impact and Industry Assessment

At $0.30 per audio hour, Saaras V4 undercuts Western cloud speech APIs by a wide margin while offering superior multilingual capabilities:

  • Google Cloud Speech-to-Text (Chirp-2): $0.96 / hour (over 3x higher cost)
  • OpenAI Whisper API: $0.36 / hour (20% higher cost, lacks native diarization)
  • AWS Transcribe: $1.44 / hour (nearly 5x higher cost)

For enterprises operating automated voice agents, outbound contact centers, or regulatory compliance auditing across India, this pricing reduction directly impacts operational viability. Running 100,000 hours of monthly customer care calls costs $30,000 on Saaras V4, compared to $96,000 on Google Cloud or $144,000 on AWS, while cutting post-processing costs associated with external diarization pipelines.

8. Summary

Sarvam AI’s release of Saaras V4 represents an important architectural shift in sovereign speech technology:

  1. Architecture: Pairs an efficient audio front-end with an in-house 3B hybrid state-space decoder to achieve linear-time inference.
  2. Multilingual Accuracy: Establishes benchmark records across 22 Indian languages while outperforming Whisper large-v3 on financial English calls.
  3. Operational Utility: Solves real-world pain points through native single-pass diarization, 8kHz telephony tuning, and five immediate text representation modes.

Developers and enterprise engineering teams can evaluate Saaras V4 today through the Sarvam AI API platform at api.sarvam.ai.