Senior Speech AI Engineer ( Human-Realistic Voices, Multilingual STT )

Senior Speech AI Engineer (Human-Realistic Voices, Multilingual STT and TTS)

Company: INTIME IT Services (INTIME Voice AI)

Location: Remote. At least 4 hours of overlap with IST.

Engagement: Full-time or part-time (minimum 20 hours per week)

Experience: 5+ years in software or ML engineering, including 2+ years in production speech AI (ASR, TTS, or voice AI)

Compensation: To be discussed, based on experience and engagement type (full-time or part-time)

About us

INTIME IT Services is an 11-year-old SAP and AI consulting company with operations in Kolkata, Dhaka, and Dubai. INTIME Voice AI is our real-time voice agent platform, live with paying clients. We are expanding across India, Bangladesh, Thailand, and the UAE. Our callers speak national languages, regional languages, and mixed-language speech, often on noisy phone lines.

The role

You make our agents understand every caller and sound indistinguishable from a human, in every main and regional language of our markets. Our benchmark is the realism of Fish Audio-class voice models: natural prosody, emotion, breathing, and conversational rhythm.

Language scope

India: Hindi, Bengali, Tamil, Telugu, Marathi, Gujarati, Kannada, Malayalam, Punjabi, Odia, Assamese, Indian English, and code-switched Hinglish

Bangladesh: Bangla (standard) and regional dialects such as Sylheti and Chittagonian, plus code-switched Bangla-English

Thailand: Central Thai and regional varieties such as Isan, Northern Thai, and Southern Thai, plus Thai-English code-switching

UAE: Gulf and Emirati Arabic, Modern Standard Arabic, English, and the main expatriate languages (Hindi, Urdu, Malayalam, Tagalog)

First launches: Bengali (India and Bangladesh), Hindi, Hinglish, Indian English, Gulf Arabic, and Thai. Regional languages follow in priority order set by client demand.

What you will do

Multilingual STT

Build evaluation sets from real 8 kHz telephony audio and 16 kHz web audio for each language and dialect

Benchmark 3 or more STT providers and open-source models per language on WER and CER, latency, streaming partial stability, and cost

Define per-language routing rules for the model routing layer

Fine-tune open-source ASR (Whisper-family, Conformer or RNN-T models) on domain and dialect data where commercial providers underperform

Handle code-switching, accents, noise, echo, and low-bitrate phone audio

Build contextual biasing and hotword boosting for names, places, brands, and domain terms

Human-realistic TTS

Benchmark 3 or more TTS providers and open-source models per language on naturalness, speaker similarity, latency, and cost

Deploy and fine-tune neural codec language model TTS and flow-matching TTS (Fish Speech/OpenAudio, CosyVoice, F5-TTS, StyleTTS 2, XTTS, or similar) for languages where commercial voices fall short

Deliver expressive control: emotion, speaking style, emphasis, speaking rate, pitch, and pause control

Generate paralinguistic features where appropriate: breaths, short fillers, laughter, and natural hesitation

Streaming TTS with time to first byte at or below 150 ms and no audible artifacts at chunk boundaries

Text normalization per language: numbers, dates, currency, abbreviations, and mixed scripts

Pronunciation lexicons and G2P handling for each language

Voice creation (authorized cloning)

Zero-shot and few-shot cloning from 10 to 60 seconds of reference audio, plus fine-tuned voices from larger datasets for premium clients

Cross-lingual voice transfer: the same voice speaking multiple languages

A consent pipeline: signed authorization, stored consent evidence, speaker verification of the voice owner, revocation, and deletion

Anti-misuse controls: no cloning of public figures or unauthorized voices, plus audio watermarking where available

Conversational realism

Semantic end-of-turn detection tuned per language, since pause patterns differ across languages

Backchannels and acknowledgments timed to the callers speech

Natural barge-in behavior: stop, listen, and respond to the interruption

Filler phrases tied to the real system state. The agent never says let me check” unless a tool is running.

Evaluation

Blind MOS and CMOS listening tests with native speakers per language

Speaker similarity scored by speaker-embedding cosine similarity (ECAPA-TDNN or WavLM-based)

TTS intelligibility measured by ASR round-trip WER

A release gate that blocks any release that drops WER, MOS, or similarity

You must have done this before

Shipped STT or TTS models or improvements to a production voice product with real users

Fine-tuned and deployed an ASR or TTS model on custom data

Built speech evaluation sets and run WER and MOS evaluations

Worked with at least one low-resource, regional, or code-switched language

Integrated 3 or more STT and 3 or more TTS systems, commercial or open-source

Built or deployed voice cloning in production

Technical requirements

Strong Python and PyTorch. torchaudio, librosa, and ESPnet, NeMo, or Hugging Face speech tooling.

Telephony audio: 8 kHz and 16 kHz, mu-law and A-law, Opus, resampling

VAD and turn-detection models

GPU training and inference, including mixed precision and quantization for low-latency serving

Strong plus

Native or fluent in one or more of: Bengali, Hindi, Thai, Arabic, Tamil, Telugu

Arabic dialect or Thai speech work

Speech-to-speech or full-duplex models

Published research or open-source speech contributions

Outcomes expected in 90 days

WER and MOS baselines for all first-launch languages on real call audio

Bengali (India and Bangladesh), Hindi, Hinglish, Indian English, Gulf Arabic, and Thai in production, using the best-performing STT and TTS per language

At least 2 regional languages production-ready

TTS naturalness of MOS 4.3 or higher in blind native-speaker tests for first-launch languages

At least 3 authorized custom voices live, with speaker similarity of 0.80 or higher and full consent records

Streaming TTS time to first byte at or below 150 ms

Please do not apply if

You have less than 2 years of production speech AI experience

Your speech work is limited to academic projects or tutorials

You have not deployed a speech model to production

How to apply

Send your CV to nayanika@intimeinc.co.in with the subject “Senior Speech AI Engineer, Full-time” or “Part-time”. Include:

The languages and dialects you have shipped

One STT or TTS improvement you delivered, with before and after WER or MOS

Samples or links to voices or models you built, if shareable

Hours per week available, if applying part-time